nautilus_persistence/writer/

feather.rs

// -------------------------------------------------------------------------------------------------
//  Copyright (C) 2015-2026 Nautech Systems Pty Ltd. All rights reserved.
//  https://nautechsystems.io
//
//  Licensed under the GNU Lesser General Public License Version 3.0 (the "License");
//  You may not use this file except in compliance with the License.
//  You may obtain a copy of the License at https://www.gnu.org/licenses/lgpl-3.0.en.html
//
//  Unless required by applicable law or agreed to in writing, software
//  distributed under the License is distributed on an "AS IS" BASIS,
//  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
//  See the License for the specific language governing permissions and
//  limitations under the License.
// -------------------------------------------------------------------------------------------------

#![expect(
    clippy::missing_errors_doc,
    reason = "Feather writer public methods forward encoding and file IO errors directly"
)]

use std::{
    any::Any,
    borrow::Cow,
    cell::RefCell,
    collections::{BTreeMap, HashMap, HashSet},
    fmt::Debug,
    fs::{self, File, OpenOptions, TryLockError},
    io::{self, BufReader, BufWriter, Seek, Write},
    path::{Path, PathBuf},
    rc::Rc,
    sync::{
        Arc,
        atomic::{AtomicU64, Ordering},
        mpsc::Sender,
    },
};

use ahash::AHashMap;
use datafusion::arrow::{
    array::{ArrayRef, DictionaryArray},
    compute::concat_batches,
    datatypes::{DataType, Field, Int32Type, Schema, SchemaRef},
    error::ArrowError,
    ipc::{reader::StreamReader, writer::StreamWriter},
    record_batch::RecordBatch,
};
use jiff::{
    SignedDuration,
    civil::Time,
    tz::{AmbiguousOffset, TimeZone},
};
use nautilus_common::{clock::Clock, live::LiveClock};
use nautilus_core::{DurationNanos, UnixNanos, time::nanos_since_unix_epoch};
use nautilus_model::{
    data::{
        Bar, CustomData, CustomDataTrait, Data, DataBatch, FundingRateUpdate, IndexPriceUpdate,
        InstrumentStatus, MarkPriceUpdate, NautilusDataType, OptionGreeks, OrderBookDelta,
        OrderBookDeltas, OrderBookDepth, QuoteTick, TradeTick,
        close::InstrumentClose,
        encode_custom_to_arrow, get_arrow_schema,
        greeks::{GreeksData, PortfolioGreeks},
    },
    events::{
        AccountState, OrderAccepted, OrderCancelRejected, OrderCanceled, OrderDenied,
        OrderEmulated, OrderEventAny, OrderExpired, OrderFillVoided, OrderFilled, OrderInitialized,
        OrderModifyRejected, OrderPendingCancel, OrderPendingUpdate, OrderRejected, OrderReleased,
        OrderSnapshot, OrderSubmitted, OrderTriggered, OrderUpdated, PositionAdjusted,
        PositionChanged, PositionClosed, PositionEvent, PositionOpened, PositionSnapshot,
    },
    instruments::{InstrumentAny, NautilusInstrumentType},
    reports::{ExecutionMassStatus, FillReport, OrderStatusReport, PositionStatusReport},
};
use nautilus_serialization::arrow::{
    EncodeToRecordBatch, KEY_TYPE_NAME, catalog_identifier_from_metadata,
    record_batch_with_identifier_column, schema_with_identifier_column,
};

use crate::{
    catalog::{
        session::DEFAULT_DATA_BATCH_CHUNK_SIZE,
        types::{
            CatalogDataType, CatalogFamily, data_path_prefix, instrument_path_prefix,
            record_path_prefix,
        },
    },
    common::custom::{
        augment_batch_with_data_type_column, schema_with_data_type_column,
        validate_custom_catalog_schema,
    },
    writer::{
        filter::{WriterRecordFilter, catalog_family},
        subscription::StreamingSinkSubscription,
        traits::{StreamingDataSink, StreamingSink},
    },
};

pub(crate) type FeatherWriteCommand =
    Box<dyn FnOnce(&mut FeatherWriter) -> Result<(), Box<dyn std::error::Error>> + Send + 'static>;

#[expect(
    clippy::needless_pass_by_value,
    reason = "map_err transfers ownership of the boxed writer error"
)]
pub(crate) fn feather_error(e: Box<dyn std::error::Error>) -> anyhow::Error {
    anyhow::anyhow!("{e}")
}

macro_rules! define_builtin_data_batch_dispatch {
    ($(($variant:ident, $type:ident, $data:ident, $batch:ident, $prefix:literal)),+ $(,)?) => {
        fn write_builtin_data_batch(
            writer: &mut FeatherWriter,
            batch: &DataBatch,
        ) -> Option<Result<(), Box<dyn std::error::Error>>> {
            match batch {
                $(
                    DataBatch::$batch(data) => Some(writer.write_batch(data.as_ref().to_vec())),
                )+
                _ => None,
            }
        }
    };
}

nautilus_model::for_each_data_type!(define_builtin_data_batch_dispatch);

pub(crate) const NAUTILUS_ARROW_METADATA_ID_COLUMN: &str = "nautilus_metadata_id";
pub(crate) const NAUTILUS_ARROW_METADATA_JSON_COLUMN: &str = "nautilus_metadata_json";

/// Arrow type of the staged metadata columns; rows of one flush share a few distinct values.
pub(crate) fn staged_metadata_data_type() -> DataType {
    DataType::Dictionary(Box::new(DataType::Int32), Box::new(DataType::Utf8))
}

pub(crate) fn staged_metadata_array<'a>(values: impl IntoIterator<Item = &'a str>) -> ArrayRef {
    Arc::new(values.into_iter().collect::<DictionaryArray<Int32Type>>())
}

#[derive(Clone, Debug, Eq, PartialEq)]
struct StagedArrowMetadataRow {
    metadata_id: String,
    metadata_json: String,
}

/// An open file and the catalog type it stages; each instrument class stages its own file.
#[derive(Debug, Clone, PartialEq, Eq, Hash)]
pub struct FileWriterPath {
    path: PathBuf,
    data_type: CatalogDataType,
}

/// Minimal `Send` time source for streaming writers.
///
/// `Live` reads the wall clock directly; `Test` reads a shared atomic that the
/// owner of the source clock (e.g. a PyO3 wrapper holding a `VirtualClock`)
/// refreshes before forwarding writer calls.
#[derive(Clone, Debug)]
pub enum WriterClock {
    /// Wall-clock time via [`nanos_since_unix_epoch`].
    Live,
    /// Externally driven time stored in a shared atomic (nanoseconds since the UNIX epoch).
    Test(Arc<AtomicU64>),
}

impl WriterClock {
    /// Returns the current timestamp in nanoseconds since the UNIX epoch.
    #[must_use]
    pub fn timestamp_ns(&self) -> UnixNanos {
        match self {
            Self::Live => UnixNanos::from(nanos_since_unix_epoch()),
            Self::Test(shared) => UnixNanos::from(shared.load(Ordering::Relaxed)),
        }
    }

    /// Builds a writer clock from a shared `dyn Clock`.
    ///
    /// Returns [`WriterClock::Live`] for a [`LiveClock`]. For any other clock
    /// (e.g. `VirtualClock`) it returns a [`WriterClock::Test`] source seeded with
    /// the clock's current time, plus the shared atomic the caller must refresh
    /// from the source clock before each forwarded writer call.
    #[must_use]
    pub fn from_shared_clock(clock: &Rc<RefCell<dyn Clock>>) -> (Self, Option<Arc<AtomicU64>>) {
        let borrowed = clock.borrow();
        let any_ref: &dyn Any = &*borrowed;
        if any_ref.downcast_ref::<LiveClock>().is_some() {
            (Self::Live, None)
        } else {
            let shared = Arc::new(AtomicU64::new(borrowed.timestamp_ns().as_u64()));
            (Self::Test(Arc::clone(&shared)), Some(shared))
        }
    }
}

fn record_batch_with_delta_staged_metadata(
    batch: &RecordBatch,
    metadata_rows: &[StagedArrowMetadataRow],
) -> anyhow::Result<RecordBatch> {
    anyhow::ensure!(
        batch.num_rows() == metadata_rows.len(),
        "Delta Feather staging metadata row count {} does not match batch row count {}",
        metadata_rows.len(),
        batch.num_rows()
    );
    anyhow::ensure!(
        batch
            .schema()
            .index_of(NAUTILUS_ARROW_METADATA_ID_COLUMN)
            .is_err(),
        "Delta Feather staging schema already has {NAUTILUS_ARROW_METADATA_ID_COLUMN}"
    );
    anyhow::ensure!(
        batch
            .schema()
            .index_of(NAUTILUS_ARROW_METADATA_JSON_COLUMN)
            .is_err(),
        "Delta Feather staging schema already has {NAUTILUS_ARROW_METADATA_JSON_COLUMN}"
    );

    let mut fields = batch
        .schema()
        .fields()
        .iter()
        .map(|field| {
            Arc::new(Field::new(
                field.name().clone(),
                field.data_type().clone(),
                field.is_nullable(),
            ))
        })
        .collect::<Vec<_>>();

    fields.push(Arc::new(Field::new(
        NAUTILUS_ARROW_METADATA_ID_COLUMN,
        staged_metadata_data_type(),
        false,
    )));
    fields.push(Arc::new(Field::new(
        NAUTILUS_ARROW_METADATA_JSON_COLUMN,
        staged_metadata_data_type(),
        false,
    )));

    let mut columns = batch.columns().to_vec();
    columns.push(staged_metadata_array(
        metadata_rows.iter().map(|row| row.metadata_id.as_str()),
    ));
    columns.push(staged_metadata_array(
        metadata_rows.iter().map(|row| row.metadata_json.as_str()),
    ));

    Ok(RecordBatch::try_new(
        Arc::new(Schema::new(fields)),
        columns,
    )?)
}

fn arrow_metadata_row(
    metadata: &HashMap<String, String>,
    fields: &arrow::datatypes::Fields,
) -> anyhow::Result<StagedArrowMetadataRow> {
    let schema_metadata = metadata
        .iter()
        .map(|(key, value)| (key.clone(), value.clone()))
        .collect::<BTreeMap<_, _>>();

    let field_metadata = fields
        .iter()
        .filter(|field| !field.metadata().is_empty())
        .map(|field| {
            (
                field.name().clone(),
                field
                    .metadata()
                    .iter()
                    .map(|(key, value)| (key.clone(), value.clone()))
                    .collect::<BTreeMap<_, _>>(),
            )
        })
        .collect::<BTreeMap<_, _>>();

    let metadata_json = serde_json::to_string(&serde_json::json!({
        "format_version": 1,
        "schema_metadata": schema_metadata,
        "field_metadata": field_metadata,
    }))?;
    let metadata_id = staged_metadata_id(&canonical_metadata_json(metadata)?);

    Ok(StagedArrowMetadataRow {
        metadata_id,
        metadata_json,
    })
}

/// Canonical schema-level metadata JSON: keys serialized in sorted order.
///
/// The staged metadata id hashes this exact string so readers can verify the schema metadata
/// independently of its storage location.
pub(crate) fn canonical_metadata_json(
    metadata: &HashMap<String, String>,
) -> anyhow::Result<String> {
    Ok(serde_json::to_string(
        &metadata
            .iter()
            .map(|(key, value)| (key.clone(), value.clone()))
            .collect::<BTreeMap<_, _>>(),
    )?)
}

pub(crate) fn staged_metadata_id(metadata_json: &str) -> String {
    format!("blake3:{}", blake3::hash(metadata_json.as_bytes()).to_hex())
}

/// File extension of a sealed Feather stream file.
pub(crate) const FEATHER_EXTENSION: &str = "feather";

/// File extension of a Feather stream file that is still being appended to.
pub(crate) const FEATHER_PARTIAL_EXTENSION: &str = "feather.partial";

/// An open local Feather file that appends Arrow IPC stream batches.
///
/// Batches go to `{name}.feather.partial`, which is renamed to `{name}.feather` when the file is
/// sealed, so readers listing `.feather` files only see complete streams. The file stays
/// exclusively locked until it is sealed, so [`recover_partial_feather_files`] can tell an open
/// file from one its writer abandoned. Written rows are
/// buffered and appended as one record batch on flush, on seal, when the buffer reaches
/// [`DEFAULT_DATA_BATCH_CHUNK_SIZE`] rows, or when size rotation may be due, because each IPC
/// batch carries a message header that outweighs a single row.
struct FeatherFile {
    writer: StreamWriter<CountingWriter<BufWriter<File>>>,
    partial_path: PathBuf,
    path: PathBuf,
    schema: SchemaRef,
    max_size: Option<u64>,
    buffer: Vec<RecordBatch>,
    buffer_rows: usize,
    buffer_bytes: u64,
}

impl FeatherFile {
    fn create(
        path: &Path,
        schema: &Schema,
        rotation_config: &RotationConfig,
    ) -> Result<Self, Box<dyn std::error::Error>> {
        let partial_path = path.with_extension(FEATHER_PARTIAL_EXTENSION);
        if let Some(parent) = path.parent() {
            fs::create_dir_all(parent)?;
        }

        let file = OpenOptions::new()
            .write(true)
            .create_new(true)
            .open(&partial_path)?;

        // Recovery can briefly lock a new, empty file before its writer does, so wait it out
        file.lock().map_err(|e| {
            format!(
                "Failed to lock Feather file {}: {e}",
                partial_path.display()
            )
        })?;

        let writer = StreamWriter::try_new(CountingWriter::new(BufWriter::new(file)), schema)?;

        let max_size = match rotation_config {
            RotationConfig::Size { max_size } => Some(*max_size),
            _ => None,
        };

        Ok(Self {
            writer,
            partial_path,
            path: path.to_path_buf(),
            schema: Arc::new(schema.clone()),
            max_size,
            buffer: Vec::new(),
            buffer_rows: 0,
            buffer_bytes: 0,
        })
    }

    // Returns true when size rotation is due
    fn write_record_batch(&mut self, batch: &RecordBatch) -> Result<bool, ArrowError> {
        let batch = if batch.schema() == self.schema {
            batch.clone()
        } else {
            RecordBatch::try_new(self.schema.clone(), batch.columns().to_vec())?
        };

        self.buffer_rows += batch.num_rows();
        self.buffer_bytes += batch.get_array_memory_size() as u64;
        self.buffer.push(batch);

        if self.buffer_rows >= DEFAULT_DATA_BATCH_CHUNK_SIZE {
            self.write_buffer()?;
        }

        let Some(max_size) = self.max_size else {
            return Ok(false);
        };

        // The in-memory size only estimates the encoded size, so write before the exact check
        if self.size() + self.buffer_bytes >= max_size {
            self.write_buffer()?;
        }

        Ok(self.size() >= max_size)
    }

    fn write_buffer(&mut self) -> Result<(), ArrowError> {
        if self.buffer.is_empty() {
            return Ok(());
        }

        let batch = concat_batches(&self.schema, &self.buffer)?;
        self.buffer.clear();
        self.buffer_rows = 0;
        self.buffer_bytes = 0;
        self.writer.write(&batch)
    }

    fn size(&self) -> u64 {
        self.writer.get_ref().bytes
    }

    fn flush(&mut self) -> Result<(), ArrowError> {
        self.write_buffer()?;
        self.writer.flush()
    }

    fn seal(mut self) -> Result<(), Box<dyn std::error::Error>> {
        self.write_buffer()?;
        self.writer.finish()?;
        let file = self
            .writer
            .into_inner()?
            .inner
            .into_inner()
            .map_err(io::IntoInnerError::into_error)?;
        file.sync_all()?;

        // Renaming before the file closes keeps it locked until it is sealed
        fs::rename(&self.partial_path, &self.path)?;
        Ok(())
    }
}

/// Counts the bytes written through the inner writer.
struct CountingWriter<W> {
    inner: W,
    bytes: u64,
}

impl<W> CountingWriter<W> {
    const fn new(inner: W) -> Self {
        Self { inner, bytes: 0 }
    }
}

impl<W: Write> Write for CountingWriter<W> {
    fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
        let written = self.inner.write(buf)?;
        self.bytes += written as u64;
        Ok(written)
    }

    fn flush(&mut self) -> io::Result<()> {
        self.inner.flush()
    }
}

/// Rotations and flushes collected by the synchronous encode path and run once it completes.
#[derive(Debug, Default)]
struct PendingIo {
    rotate_paths: Vec<FileWriterPath>,
    flush_due: bool,
}

/// Configuration for file rotation.
#[derive(Debug, Clone)]
pub enum RotationConfig {
    /// Rotate based on file size.
    Size {
        /// Maximum buffer size in bytes before rotation.
        max_size: u64,
    },
    /// Rotate based on a time interval.
    Interval {
        /// Interval in nanoseconds.
        interval_ns: u64,
    },
    /// Rotate based on scheduled dates.
    ScheduledDates {
        /// Interval in nanoseconds.
        interval_ns: u64,
        /// Time of day for rotation (nanoseconds since midnight).
        rotation_time: UnixNanos,
        /// Timezone for rotation calculations.
        rotation_timezone: TimeZone,
    },
    /// No automatic rotation.
    NoRotation,
}

/// Streams encoded data into one local Feather file per data or record type, and per class for
/// instruments.
///
/// Rows of every identifier share their type's file. Each row carries its catalog `identifier`
/// and the Arrow metadata it was encoded with, such as its instrument's precision, so readers
/// restore per-row metadata the file schema cannot hold.
///
/// The [`Self::write`] method accepts a data value of generic type `T`, encodes it using `T`'s
/// metadata via `EncodeToRecordBatch`, routes it by `CatalogFamily`, and appends it to that key's
/// open file.
///
/// # File Lifecycle
///
/// Open files are tracked by their final `.feather` paths, while writes go to `.feather.partial`
/// files. Sealing completes the stream and renames the partial file to its final path. A file is
/// sealed and replaced only on rotation, [`Self::seal`], or [`Self::close`].
///
/// # Flushing
///
/// [`Self::flush`] pushes buffered bytes to disk without starting a new file. `flush_interval_ms`
/// controls automatic flushing in milliseconds; `0` disables automatic flushing.
pub struct FeatherWriter {
    directory: PathBuf,
    clock: WriterClock,
    rotation_config: RotationConfig,
    included_types: Option<HashSet<CatalogDataType>>,
    record_filter: Option<WriterRecordFilter>,
    writers: HashMap<FileWriterPath, FeatherFile>,
    path_sequences: HashMap<CatalogDataType, (UnixNanos, u64)>,
    sealed_tx: Option<Sender<PathBuf>>,
    next_rotation_times: HashMap<FileWriterPath, UnixNanos>,
    flush_interval_ms: u64,
    last_flush_ns: UnixNanos,
    pending_write_error: Option<String>,
}

impl FeatherWriter {
    /// Creates a new [`FeatherWriter`] instance.
    #[must_use]
    pub fn new(
        directory: PathBuf,
        clock: WriterClock,
        rotation_config: RotationConfig,
        included_types: Option<HashSet<CatalogDataType>>,
        flush_interval_ms: Option<u64>,
    ) -> Self {
        let flush_interval_ms = flush_interval_ms.unwrap_or(1000); // Default 1 second
        let last_flush_ns = clock.timestamp_ns();

        Self {
            directory,
            clock,
            rotation_config,
            included_types,
            record_filter: None,
            writers: HashMap::new(),
            path_sequences: HashMap::new(),
            sealed_tx: None,
            next_rotation_times: HashMap::new(),
            flush_interval_ms,
            last_flush_ns,
            pending_write_error: None,
        }
    }

    /// Sets typed record-family filter for subsequent writes.
    #[must_use]
    pub fn with_record_filter(mut self, record_filter: Option<WriterRecordFilter>) -> Self {
        self.record_filter = record_filter;
        self
    }

    pub(crate) fn with_sealed_sender(mut self, sealed_tx: Sender<PathBuf>) -> Self {
        self.sealed_tx = Some(sealed_tx);
        self
    }

    /// Writes a single data value.
    ///
    /// This is the user entry point. The data is encoded into a `RecordBatch` and appended to the
    /// open file for its type, or for its class when the data is an instrument.
    pub fn write<T>(&mut self, data: T) -> Result<(), Box<dyn std::error::Error>>
    where
        T: EncodeToRecordBatch + CatalogFamily + 'static,
    {
        let metadata = T::metadata(&data);
        let identifier = catalog_identifier_from_metadata(&metadata);
        let data_type = Self::staged_data_type::<T>(&metadata)?;

        if !self.should_write_record(&data_type, identifier.as_deref()) {
            return Ok(());
        }

        let path = self.get_writer_path(data_type);

        // Create a new FileWriter if one does not exist.
        if !self.writers.contains_key(&path) {
            self.create_writer::<T>(path.clone(), &data)?;
        }

        // Encode the data into a RecordBatch using T's encoding logic.
        let batch = T::encode_batch(&metadata, &[data])?;
        let metadata_row = arrow_metadata_row(&metadata, batch.schema().fields())?;
        let batch = record_batch_with_identifier_column(batch, identifier.as_deref())?;
        let batch =
            record_batch_with_delta_staged_metadata(&batch, std::slice::from_ref(&metadata_row))?;

        // Write the RecordBatch to the appropriate FileWriter.
        let mut pending = PendingIo::default();

        self.stage_batch_write(path, &batch, &mut pending)?;
        pending.flush_due = self.flush_is_due();

        self.complete_pending_io(&pending)
    }

    /// Writes a batch of data values as one or more `RecordBatch`es.
    ///
    /// Uses `T::chunk_metadata` to derive the file schema metadata. This protects
    /// types like `OrderBookDelta` from having their file metadata poisoned by a
    /// leading sentinel row (e.g. `BookAction::Clear`, which carries
    /// `price_precision=0, size_precision=0`).
    ///
    /// Rows are grouped by identifier so each group encodes with its own metadata before it is
    /// appended to the type's file.
    pub fn write_batch<T>(&mut self, data: Vec<T>) -> Result<(), Box<dyn std::error::Error>>
    where
        T: EncodeToRecordBatch + CatalogFamily + 'static,
    {
        if data.is_empty() || !self.should_write_family(&T::catalog_family()) {
            return Ok(());
        }

        let mut groups: AHashMap<Option<String>, (CatalogDataType, Vec<T>)> = AHashMap::new();

        for item in data {
            let metadata = T::metadata(&item);
            let identifier = catalog_identifier_from_metadata(&metadata);
            let data_type = Self::staged_data_type::<T>(&metadata)?;

            if !self.should_write_record(&data_type, identifier.as_deref()) {
                continue;
            }

            groups
                .entry(identifier)
                .or_insert_with(|| (data_type, Vec::new()))
                .1
                .push(item);
        }

        if groups.is_empty() {
            return Ok(());
        }

        let mut pending = PendingIo::default();

        for (data_type, group) in groups.into_values() {
            let path = self.get_writer_path(data_type);
            let metadata = T::chunk_metadata(&group);

            if !self.writers.contains_key(&path) {
                self.create_writer_with_metadata::<T>(path.clone(), metadata.clone())?;
            }

            let identifier = catalog_identifier_from_metadata(&metadata);
            let batch = T::encode_batch(&metadata, &group)?;
            let metadata_rows = group
                .iter()
                .map(T::metadata)
                .map(|metadata| arrow_metadata_row(&metadata, batch.schema().fields()))
                .collect::<anyhow::Result<Vec<_>>>()?;
            let batch = record_batch_with_identifier_column(batch, identifier.as_deref())?;
            let batch = record_batch_with_delta_staged_metadata(&batch, &metadata_rows)?;

            self.stage_batch_write(path, &batch, &mut pending)?;
        }

        pending.flush_due = self.flush_is_due();

        self.complete_pending_io(&pending)
    }

    // Instrument classes have different fields, so each class stages its own file
    fn staged_data_type<T: CatalogFamily>(
        metadata: &HashMap<String, String>,
    ) -> Result<CatalogDataType, Box<dyn std::error::Error>> {
        let family = T::catalog_family();
        if family != CatalogDataType::Data(NautilusDataType::Instrument) {
            return Ok(family);
        }

        let class = metadata
            .get(KEY_TYPE_NAME)
            .ok_or("Instrument metadata has no type_name")?
            .parse::<NautilusInstrumentType>()?;
        Ok(CatalogDataType::Instrument(class))
    }

    fn stage_batch_write(
        &mut self,
        path: FileWriterPath,
        batch: &RecordBatch,
        pending: &mut PendingIo,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let Some(file) = self.writers.get_mut(&path) else {
            return Ok(());
        };

        let should_rotate = match file.write_record_batch(batch) {
            Ok(should_rotate) => should_rotate,
            Err(e) => {
                self.abandon_file(&path);
                return Err(e.into());
            }
        };

        if should_rotate || self.check_scheduled_rotation(&path) {
            pending.rotate_paths.push(path);
        }

        Ok(())
    }

    /// Returns whether the auto-flush interval has elapsed since the last flush.
    fn flush_is_due(&self) -> bool {
        if self.flush_interval_ms == 0 {
            return false; // Auto-flush disabled
        }

        let now_ns = self.clock.timestamp_ns();
        let elapsed_ms = now_ns.as_u64().saturating_sub(self.last_flush_ns.as_u64()) / 1_000_000;
        elapsed_ms >= self.flush_interval_ms
    }

    fn complete_pending_io(
        &mut self,
        pending: &PendingIo,
    ) -> Result<(), Box<dyn std::error::Error>> {
        for path in &pending.rotate_paths {
            self.seal_file(path)?;
        }

        if pending.flush_due {
            self.flush()?;
        }

        Ok(())
    }

    fn check_scheduled_rotation(&mut self, path: &FileWriterPath) -> bool {
        match &self.rotation_config {
            RotationConfig::Interval { interval_ns } => {
                let now = self.clock.timestamp_ns();
                let next_rotation = self.next_rotation_times.get(path).copied();

                match next_rotation {
                    None => {
                        self.next_rotation_times
                            .insert(path.clone(), now + DurationNanos::new(*interval_ns));
                        false
                    }
                    Some(next) if now >= next => {
                        self.next_rotation_times
                            .insert(path.clone(), now + DurationNanos::new(*interval_ns));
                        true
                    }
                    _ => false,
                }
            }
            RotationConfig::ScheduledDates {
                interval_ns,
                rotation_time,
                rotation_timezone,
            } => {
                let now = self.clock.timestamp_ns();
                let next_rotation = self.next_rotation_times.get(path).copied();

                match next_rotation {
                    None => {
                        let next = self.calculate_next_scheduled_rotation(
                            *rotation_time,
                            rotation_timezone,
                            *interval_ns,
                        );
                        self.next_rotation_times.insert(path.clone(), next);
                        false
                    }
                    Some(next) if now >= next => {
                        let next = self.calculate_next_scheduled_rotation(
                            *rotation_time,
                            rotation_timezone,
                            *interval_ns,
                        );
                        self.next_rotation_times.insert(path.clone(), next);
                        true
                    }
                    _ => false,
                }
            }
            _ => false,
        }
    }

    fn calculate_next_scheduled_rotation(
        &self,
        rotation_time: UnixNanos,
        rotation_timezone: &TimeZone,
        interval_ns: u64,
    ) -> UnixNanos {
        let now_utc = self.clock.timestamp_ns().to_datetime_utc();
        let now_local = rotation_timezone.to_datetime(now_utc);

        let rotation_time_secs = u32::try_from(*rotation_time / 1_000_000_000).unwrap_or(0);
        let rotation_time_nanos = i32::try_from(*rotation_time % 1_000_000_000).unwrap_or(0);

        let rotation_time = if rotation_time_secs < 86_400 {
            Time::new(
                i8::try_from(rotation_time_secs / 3_600).unwrap_or(0),
                i8::try_from(rotation_time_secs % 3_600 / 60).unwrap_or(0),
                i8::try_from(rotation_time_secs % 60).unwrap_or(0),
                rotation_time_nanos,
            )
            .unwrap_or(Time::MIN)
        } else {
            Time::MIN
        };

        let local_rotation = now_local.date().to_datetime(rotation_time);
        let ambiguous = rotation_timezone.to_ambiguous_timestamp(local_rotation);

        let mut next_rotation = match ambiguous.offset() {
            AmbiguousOffset::Gap { .. } => now_utc,
            _ => ambiguous.earlier().unwrap_or(now_utc),
        };

        if next_rotation <= now_utc {
            // If the time has already passed today, we would usually add the interval
            // But let's align exactly with how Python does it:
            while next_rotation <= now_utc {
                next_rotation += SignedDuration::from_nanos_i128(i128::from(interval_ns));
            }
        }

        UnixNanos::from(u64::try_from(next_rotation.as_nanosecond()).unwrap_or(0))
    }

    // Seals the open file at `path`; the next write for its key opens a new file. A failed seal
    // leaves the stream unfinished, so the file is recovered to its complete batches instead.
    fn seal_file(&mut self, path: &FileWriterPath) -> Result<(), Box<dyn std::error::Error>> {
        self.next_rotation_times.remove(path);

        let Some(file) = self.writers.remove(path) else {
            return Ok(());
        };

        let partial_path = file.partial_path.clone();
        let result = file
            .seal()
            .inspect_err(|_| recover_released_feather_file(&partial_path));
        self.notify_sealed(path);
        result
    }

    // A failed append can leave a partial IPC message, so the file is recovered to its complete
    // batches rather than sealed as written; the next write for its key opens a new file.
    fn abandon_file(&mut self, path: &FileWriterPath) {
        self.next_rotation_times.remove(path);

        let Some(file) = self.writers.remove(path) else {
            return;
        };

        let partial_path = file.partial_path.clone();
        drop(file);

        log::error!(
            "Abandoned Feather file {} after a failed write",
            path.path.display()
        );

        recover_released_feather_file(&partial_path);
        self.notify_sealed(path);
    }

    fn notify_sealed(&self, path: &FileWriterPath) {
        if let Some(sealed_tx) = &self.sealed_tx
            && path.path.exists()
        {
            let _ = sealed_tx.send(path.path.clone());
        }
    }

    /// Creates (and inserts) a new `FileWriter` for type T.
    fn create_writer<T>(
        &mut self,
        path: FileWriterPath,
        data: &T,
    ) -> Result<(), Box<dyn std::error::Error>>
    where
        T: EncodeToRecordBatch + CatalogFamily + 'static,
    {
        self.create_writer_with_metadata::<T>(path, T::metadata(data))
    }

    /// Creates (and inserts) a new `FileWriter` for type T with pre-computed metadata.
    ///
    /// Use this variant when the caller has selected metadata from a chunk
    /// (e.g. via `T::chunk_metadata`) to avoid schema poisoning by sentinel rows.
    fn create_writer_with_metadata<T>(
        &mut self,
        path: FileWriterPath,
        metadata: HashMap<String, String>,
    ) -> Result<(), Box<dyn std::error::Error>>
    where
        T: EncodeToRecordBatch + CatalogFamily + 'static,
    {
        let schema = Self::schema_with_delta_staging_columns(&schema_with_identifier_column(
            &T::get_schema(Some(metadata)),
        ));

        let file = FeatherFile::create(&path.path, &schema, &self.rotation_config)?;
        self.writers.insert(path, file);
        Ok(())
    }

    /// Creates (and inserts) a new `FeatherBuffer` for custom data at the given path.
    fn create_custom_writer(
        &mut self,
        path: FileWriterPath,
        type_name: &str,
    ) -> Result<(), Box<dyn std::error::Error>> {
        if self.writers.contains_key(&path) {
            return Ok(());
        }

        let base_schema = get_arrow_schema(type_name).ok_or_else(|| {
            format!("Custom data type \"{type_name}\" is not registered for Arrow encoding")
        })?;

        let schema = schema_with_data_type_column(base_schema.as_ref(), type_name);

        let schema =
            Self::schema_with_delta_staging_columns(&schema_with_identifier_column(&schema));

        let file = FeatherFile::create(&path.path, &schema, &self.rotation_config)
            .map_err(|e| format!("Failed to create Feather file for custom {type_name}: {e}"))?;
        self.writers.insert(path, file);
        Ok(())
    }

    /// Encodes a single `CustomData` into a `RecordBatch` with `data_type` column (catalog-compatible).
    pub(crate) fn encode_custom_to_batch(
        custom: &CustomData,
    ) -> Result<RecordBatch, Box<dyn std::error::Error>> {
        let type_name = custom.data.type_name();
        let data_type_json = custom
            .data_type
            .to_persistence_json()
            .map_err(|e| format!("Failed to serialize data_type for persistence: {e}"))?;
        let dt_meta = custom.data_type.metadata_string_map();
        let items: [Arc<dyn CustomDataTrait>; 1] = [Arc::clone(&custom.data)];

        let batch = encode_custom_to_arrow(type_name, &items)
            .map_err(|e| format!("Failed to encode custom data: {e}"))?
            .ok_or_else(|| {
                format!("Custom data type \"{type_name}\" is not registered for Arrow")
            })?;

        let batch = augment_batch_with_data_type_column(
            &batch,
            &data_type_json,
            type_name,
            dt_meta.as_ref(),
        )
        .map_err(|e| e.to_string())?;
        Ok(batch)
    }

    fn schema_with_delta_staging_columns(schema: &Schema) -> Schema {
        let mut fields = schema
            .fields()
            .iter()
            .map(|field| {
                Arc::new(Field::new(
                    field.name().clone(),
                    field.data_type().clone(),
                    field.is_nullable(),
                ))
            })
            .collect::<Vec<_>>();

        if schema.index_of(NAUTILUS_ARROW_METADATA_ID_COLUMN).is_err() {
            fields.push(Arc::new(Field::new(
                NAUTILUS_ARROW_METADATA_ID_COLUMN,
                staged_metadata_data_type(),
                false,
            )));
        }

        if schema
            .index_of(NAUTILUS_ARROW_METADATA_JSON_COLUMN)
            .is_err()
        {
            fields.push(Arc::new(Field::new(
                NAUTILUS_ARROW_METADATA_JSON_COLUMN,
                staged_metadata_data_type(),
                false,
            )));
        }

        Schema::new(fields)
    }

    /// Flushes buffered bytes of every open file to disk.
    ///
    /// This is called automatically based on `flush_interval_ms` if configured, but can also
    /// be called manually by the client. Files stay open, so flushing never starts a new file.
    pub fn flush(&mut self) -> Result<(), Box<dyn std::error::Error>> {
        let paths = self.writers.keys().cloned().collect::<Vec<_>>();

        for path in paths {
            if let Some(file) = self.writers.get_mut(&path)
                && let Err(e) = file.flush()
            {
                self.abandon_file(&path);
                return Err(e.into());
            }
        }

        self.last_flush_ns = self.clock.timestamp_ns();

        if let Some(error) = self.pending_write_error.take() {
            return Err(error.into());
        }

        Ok(())
    }

    /// Seals every open file so its complete stream is visible as a `.feather` file.
    ///
    /// Writes after sealing open new files, so this is the boundary promotion uses.
    pub fn seal(&mut self) -> Result<(), Box<dyn std::error::Error>> {
        let paths = self.writers.keys().cloned().collect::<Vec<_>>();
        let mut first_error = None;

        for path in paths {
            if let Err(e) = self.seal_file(&path) {
                log::error!("Failed to seal Feather file {}: {e}", path.path.display());
                first_error.get_or_insert(e);
            }
        }

        match first_error {
            Some(e) => Err(e),
            None => Ok(()),
        }
    }

    /// Seals all open files and reports any write error recorded since the last flush.
    ///
    /// Writes after closing open new files.
    pub fn close(&mut self) -> Result<(), Box<dyn std::error::Error>> {
        self.seal()?;

        if let Some(error) = self.pending_write_error.take() {
            return Err(error.into());
        }

        Ok(())
    }

    /// Returns whether the writer has been closed (all writers cleared).
    #[must_use]
    pub fn is_closed(&self) -> bool {
        self.writers.is_empty()
    }

    /// Returns the current file size and path of each open file, keyed by the type it stages.
    #[must_use]
    pub fn get_current_file_info(&self) -> HashMap<CatalogDataType, (u64, String)> {
        self.writers
            .iter()
            .map(|(path, file)| {
                (
                    path.data_type.clone(),
                    (file.size(), path.path.display().to_string()),
                )
            })
            .collect()
    }

    /// Returns the next rotation time of the file staging `data_type`, if set.
    ///
    /// Each instrument class stages its own file, so pass [`CatalogDataType::Instrument`].
    #[must_use]
    pub fn get_next_rotation_time(&self, data_type: &CatalogDataType) -> Option<UnixNanos> {
        self.next_rotation_times
            .iter()
            .find(|(path, _)| path.data_type == *data_type)
            .map(|(_, &next)| next)
    }

    /// Determines whether a family can be written before row-level identifier checks.
    fn should_write_family(&self, family: &CatalogDataType) -> bool {
        self.included_types
            .as_ref()
            .is_none_or(|included| included.contains(family))
            && self
                .record_filter
                .as_ref()
                .is_none_or(|filter| filter.contains(family))
    }

    fn should_write_record(&self, data_type: &CatalogDataType, identifier: Option<&str>) -> bool {
        self.included_types
            .as_ref()
            .is_none_or(|included| included.contains(&catalog_family(data_type)))
            && self
                .record_filter
                .as_ref()
                .is_none_or(|filter| filter.allows(data_type, identifier))
    }

    fn reserve_writer_path(&mut self, data_type: CatalogDataType) -> FileWriterPath {
        let now = self.clock.timestamp_ns();
        let (timestamp, first_sequence) = self
            .path_sequences
            .get(&data_type)
            .copied()
            .filter(|(timestamp, _)| *timestamp >= now)
            .unwrap_or((now, 0));

        for sequence in first_sequence.. {
            let path = self.build_writer_path(&data_type, timestamp, sequence);

            // Files left by an earlier writer in the same directory keep their names
            if path.exists() || path.with_extension(FEATHER_PARTIAL_EXTENSION).exists() {
                continue;
            }

            self.path_sequences
                .insert(data_type.clone(), (timestamp, sequence + 1));
            return FileWriterPath { path, data_type };
        }

        unreachable!("unbounded writer path sequence exhausted")
    }

    fn build_writer_path(
        &self,
        data_type: &CatalogDataType,
        timestamp: UnixNanos,
        sequence: u64,
    ) -> PathBuf {
        let (directory, file_stem) = match data_type {
            CatalogDataType::Data(NautilusDataType::Custom { type_name }) => (
                self.directory.join("data").join("custom").join(type_name),
                Cow::Borrowed(type_name.as_str()),
            ),
            CatalogDataType::Data(data_type) => {
                let prefix = data_path_prefix(data_type);
                (self.directory.join(prefix.as_ref()), prefix)
            }
            CatalogDataType::Record(record_type) => {
                let prefix = record_path_prefix(record_type);
                (self.directory.join(prefix.as_ref()), prefix)
            }
            CatalogDataType::Instrument(class) => {
                let prefix = data_path_prefix(&NautilusDataType::Instrument);
                (
                    self.directory
                        .join(prefix.as_ref())
                        .join(instrument_path_prefix(class)),
                    prefix,
                )
            }
        };

        directory.join(Self::timestamped_feather_file_name(
            &file_stem, timestamp, sequence,
        ))
    }

    fn timestamped_feather_file_name(stem: &str, timestamp: UnixNanos, sequence: u64) -> String {
        if sequence == 0 {
            format!("{stem}_{timestamp}.{FEATHER_EXTENSION}")
        } else {
            format!("{stem}_{timestamp}-{sequence}.{FEATHER_EXTENSION}")
        }
    }

    /// Returns the open file of a custom type, or reserves `data/custom/{type_name}/` for it.
    fn get_writer_path_custom(
        &mut self,
        type_name: &str,
    ) -> Result<FileWriterPath, Box<dyn std::error::Error>> {
        let data_type = CatalogDataType::from(NautilusDataType::Custom {
            type_name: type_name.to_string(),
        });

        if let Some(existing) = self.writers.keys().find(|path| path.data_type == data_type) {
            return Ok(existing.clone());
        }

        if let Some(schema) = get_arrow_schema(type_name) {
            validate_custom_catalog_schema(type_name, &schema)?;
        }

        Ok(self.reserve_writer_path(data_type))
    }

    /// Returns the open file staging `data_type`, or reserves a new one.
    fn get_writer_path(&mut self, data_type: CatalogDataType) -> FileWriterPath {
        if let Some(existing) = self.writers.keys().find(|path| path.data_type == data_type) {
            return existing.clone();
        }

        self.reserve_writer_path(data_type)
    }

    /// Writes a Data enum value to the appropriate writer.
    ///
    /// This is a convenience method that routes the Data enum to the appropriate
    /// typed write method.
    pub fn write_data(&mut self, data: Data) -> Result<(), Box<dyn std::error::Error>> {
        match data {
            Data::Instrument(instrument) => self.write(*instrument),
            Data::Quote(quote) => self.write(quote),
            Data::Trade(trade) => self.write(trade),
            Data::Bar(bar) => self.write(bar),
            Data::BookDelta(delta) => self.write(delta),
            Data::BookDepth(depth) => self.write(*depth),
            Data::IndexPrice(price) => self.write(price),
            Data::MarkPrice(price) => self.write(price),
            Data::FundingRate(funding) => self.write(funding),
            Data::InstrumentStatus(status) => self.write(status),
            Data::OptionGreeks(greeks) => self.write(greeks),
            Data::InstrumentClose(close) => self.write(close),
            Data::Custom(custom) => self.write_custom_data(&custom),
            Data::BookDeltas(deltas_api) => {
                // Batch write so chunk_metadata can skip a leading BookAction::Clear sentinel
                self.write_batch(deltas_api.deltas.clone())
            }
            #[cfg(feature = "defi")]
            Data::Defi(_) => Err("Unsupported DeFi data variant for feather writes".into()),
            #[cfg(not(feature = "defi"))]
            #[allow(
                unreachable_patterns,
                reason = "DeFi variants can exist without this crate's defi feature"
            )]
            _ => Err("Unsupported DeFi data variant for feather writes".into()),
        }
    }

    /// Writes mixed data as typed Arrow batches.
    ///
    /// # Panics
    ///
    /// Panics if the built-in data batch dispatch becomes non-exhaustive.
    #[expect(
        clippy::needless_pass_by_value,
        reason = "the public writer API accepts ownership of each submitted data batch"
    )]
    pub fn write_data_batch(&mut self, data: Vec<Data>) -> Result<(), Box<dyn std::error::Error>> {
        for batch in DataBatch::from_data_vec_grouped(&data)? {
            match &batch {
                DataBatch::Custom(data) => {
                    for custom in data.as_ref() {
                        self.write_custom_data(custom)?;
                    }
                }
                batch => write_builtin_data_batch(self, batch)
                    .expect("built-in data batch dispatch is exhaustive")?,
            }
        }

        Ok(())
    }

    /// Writes supported message bus value.
    pub fn write_any_message(
        &mut self,
        message: &dyn Any,
    ) -> Result<bool, Box<dyn std::error::Error>> {
        let Some(command) = Self::write_command(message) else {
            return Ok(false);
        };

        if let Err(e) = command(self) {
            self.record_write_error(e.to_string());
            return Err(e);
        }

        Ok(true)
    }

    pub(crate) fn record_write_error(&mut self, error: String) {
        self.pending_write_error.get_or_insert(error);
    }

    pub(crate) fn write_command(message: &dyn Any) -> Option<FeatherWriteCommand> {
        macro_rules! try_write {
            ($message:expr, $type:ty) => {
                if let Some(value) = $message.downcast_ref::<$type>() {
                    let value = value.clone();
                    return Some(Box::new(move |writer| writer.write(value)));
                }
            };
        }

        try_write!(message, QuoteTick);
        try_write!(message, TradeTick);
        try_write!(message, Bar);
        try_write!(message, OrderBookDelta);
        try_write!(message, OrderBookDepth);
        try_write!(message, IndexPriceUpdate);
        try_write!(message, MarkPriceUpdate);
        try_write!(message, FundingRateUpdate);
        try_write!(message, InstrumentStatus);
        try_write!(message, OptionGreeks);
        try_write!(message, InstrumentClose);
        try_write!(message, InstrumentAny);
        try_write!(message, AccountState);
        try_write!(message, OrderInitialized);
        try_write!(message, OrderDenied);
        try_write!(message, OrderEmulated);
        try_write!(message, OrderSubmitted);
        try_write!(message, OrderAccepted);
        try_write!(message, OrderRejected);
        try_write!(message, OrderPendingCancel);
        try_write!(message, OrderCanceled);
        try_write!(message, OrderCancelRejected);
        try_write!(message, OrderExpired);
        try_write!(message, OrderTriggered);
        try_write!(message, OrderPendingUpdate);
        try_write!(message, OrderReleased);
        try_write!(message, OrderModifyRejected);
        try_write!(message, OrderUpdated);
        try_write!(message, OrderFilled);
        try_write!(message, OrderFillVoided);
        try_write!(message, PositionOpened);
        try_write!(message, PositionChanged);
        try_write!(message, PositionClosed);
        try_write!(message, PositionAdjusted);
        try_write!(message, OrderSnapshot);
        try_write!(message, PositionSnapshot);
        try_write!(message, OrderStatusReport);
        try_write!(message, FillReport);
        try_write!(message, PositionStatusReport);
        try_write!(message, ExecutionMassStatus);
        try_write!(message, GreeksData);
        try_write!(message, PortfolioGreeks);

        if let Some(deltas) = message.downcast_ref::<OrderBookDeltas>() {
            let deltas = deltas.deltas.clone();
            return Some(Box::new(move |writer| writer.write_batch(deltas)));
        }

        if let Some(data) = message.downcast_ref::<Data>() {
            let data = data.clone();
            return Some(Box::new(move |writer| writer.write_data(data)));
        }

        if let Some(custom) = message.downcast_ref::<CustomData>() {
            let custom = custom.clone();
            return Some(Box::new(move |writer| {
                writer.write_data(Data::Custom(custom))
            }));
        }

        if let Some(event) = message.downcast_ref::<OrderEventAny>() {
            let event = event.clone();
            return Some(Box::new(move |writer| writer.write_order_event(&event)));
        }

        if let Some(event) = message.downcast_ref::<PositionEvent>() {
            let event = event.clone();
            return Some(Box::new(move |writer| writer.write_position_event(&event)));
        }

        None
    }

    fn write_order_event(
        &mut self,
        event: &OrderEventAny,
    ) -> Result<(), Box<dyn std::error::Error>> {
        match event {
            OrderEventAny::Initialized(event) => self.write(event.clone()),
            OrderEventAny::Denied(event) => self.write(*event),
            OrderEventAny::Emulated(event) => self.write(*event),
            OrderEventAny::Released(event) => self.write(*event),
            OrderEventAny::Submitted(event) => self.write(*event),
            OrderEventAny::Accepted(event) => self.write(*event),
            OrderEventAny::Rejected(event) => self.write(*event),
            OrderEventAny::Canceled(event) => self.write(*event),
            OrderEventAny::Expired(event) => self.write(*event),
            OrderEventAny::Triggered(event) => self.write(*event),
            OrderEventAny::PendingUpdate(event) => self.write(*event),
            OrderEventAny::PendingCancel(event) => self.write(*event),
            OrderEventAny::ModifyRejected(event) => self.write(*event),
            OrderEventAny::CancelRejected(event) => self.write(*event),
            OrderEventAny::Updated(event) => self.write(*event),
            OrderEventAny::Filled(event) => self.write(event.clone()),
            OrderEventAny::FillVoided(event) => self.write(event.clone()),
        }
    }

    fn write_position_event(
        &mut self,
        event: &PositionEvent,
    ) -> Result<(), Box<dyn std::error::Error>> {
        match event {
            PositionEvent::PositionOpened(event) => self.write(event.clone()),
            PositionEvent::PositionChanged(event) => self.write(event.clone()),
            PositionEvent::PositionClosed(event) => self.write(event.clone()),
            PositionEvent::PositionAdjusted(event) => self.write(*event),
        }
    }

    /// Writes a single custom data value (catalog layout: `data/custom/{type_name}[/{identifier}]`).
    fn write_custom_data(&mut self, custom: &CustomData) -> Result<(), Box<dyn std::error::Error>> {
        let batch = Self::encode_custom_to_batch(custom)?;
        self.write_custom_batch(custom, batch)
    }

    pub(crate) fn write_custom_batch(
        &mut self,
        custom: &CustomData,
        mut batch: RecordBatch,
    ) -> Result<(), Box<dyn std::error::Error>> {
        let type_name = custom.data.type_name();
        let identifier = custom.data_type.identifier().map(String::from);

        if !self.should_write_custom(type_name, identifier.as_deref()) {
            return Ok(());
        }

        let path = self.get_writer_path_custom(type_name)?;
        if !self.writers.contains_key(&path) {
            self.create_custom_writer(path.clone(), type_name)?;
        }

        let metadata_row = arrow_metadata_row(batch.schema().metadata(), batch.schema().fields())?;
        batch = record_batch_with_identifier_column(batch, custom.data_type.identifier())?;
        batch =
            record_batch_with_delta_staged_metadata(&batch, std::slice::from_ref(&metadata_row))?;

        let mut pending = PendingIo::default();

        self.stage_batch_write(path, &batch, &mut pending)?;
        pending.flush_due = self.flush_is_due();

        self.complete_pending_io(&pending)
    }

    pub(crate) fn should_write_custom(&self, type_name: &str, identifier: Option<&str>) -> bool {
        self.should_write_record(
            &CatalogDataType::from(NautilusDataType::Custom {
                type_name: type_name.to_string(),
            }),
            identifier,
        )
    }

    /// Writes an instrument to the appropriate writer.
    ///
    /// Each instrument class stages its own file, and every row carries its instrument ID.
    pub fn write_instrument(
        &mut self,
        instrument: InstrumentAny,
    ) -> Result<(), Box<dyn std::error::Error>> {
        self.write(instrument)
    }

    /// Subscribes to all messages on the message bus (pattern "*").
    ///
    /// This will automatically write all supported data types that are published
    /// on the message bus to the feather files.
    ///
    /// The writer must be wrapped in `Rc<RefCell<>>` to be shareable with the message bus handler.
    ///
    /// Note: Writes are synchronous local file appends.
    pub fn subscribe_to_message_bus(
        writer: Rc<RefCell<Self>>,
    ) -> Result<StreamingSinkSubscription, Box<dyn std::error::Error>> {
        let sink: StreamingDataSink = Box::new(writer);
        Ok(StreamingSinkSubscription::subscribe(
            Rc::new(RefCell::new(sink)),
            None,
        ))
    }

    /// Unsubscribes from message bus.
    pub fn unsubscribe_from_message_bus(handler: &StreamingSinkSubscription) {
        handler.unsubscribe();
    }
}

impl Debug for FeatherWriter {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct(stringify!(FeatherWriter))
            .finish_non_exhaustive()
    }
}

/// Drop contract: seals open files so their streams stay readable and promotable.
///
/// Errors are logged, never raised. Call [`FeatherWriter::close`] for deterministic sealing.
impl Drop for FeatherWriter {
    fn drop(&mut self) {
        if !self.writers.is_empty() {
            let _ = self.seal();
        }
    }
}

impl StreamingSink for FeatherWriter {
    fn write_data(&mut self, data: Data) -> anyhow::Result<()> {
        Self::write_data(self, data).map_err(feather_error)
    }

    fn write_batch(&mut self, data: Vec<Data>) -> anyhow::Result<()> {
        Self::write_data_batch(self, data).map_err(feather_error)
    }

    fn write_any(&mut self, message: &dyn Any) -> anyhow::Result<bool> {
        self.write_any_message(message).map_err(feather_error)
    }

    fn flush(&mut self) -> anyhow::Result<()> {
        Self::flush(self).map_err(feather_error)
    }

    fn close(&mut self) -> anyhow::Result<()> {
        Self::close(self).map_err(feather_error)
    }
}

impl StreamingSink for Rc<RefCell<FeatherWriter>> {
    fn write_data(&mut self, data: Data) -> anyhow::Result<()> {
        StreamingSink::write_data(&mut *self.borrow_mut(), data)
    }

    fn write_batch(&mut self, data: Vec<Data>) -> anyhow::Result<()> {
        StreamingSink::write_batch(&mut *self.borrow_mut(), data)
    }

    fn write_any(&mut self, message: &dyn Any) -> anyhow::Result<bool> {
        StreamingSink::write_any(&mut *self.borrow_mut(), message)
    }

    fn flush(&mut self) -> anyhow::Result<()> {
        StreamingSink::flush(&mut *self.borrow_mut())
    }

    fn close(&mut self) -> anyhow::Result<()> {
        StreamingSink::close(&mut *self.borrow_mut())
    }
}

/// Seals every abandoned `.feather.partial` file under `directory`.
///
/// A writer locks each file until it seals it, so a file that recovery can lock was left by a
/// writer that crashed, was killed, or released it after a failed write or seal. Recovery keeps the
/// complete record batches, drops any bytes after them with a warning, and renames the file to
/// `.feather`. It removes a file with no complete record batch and leaves an empty one, which a new
/// writer may not have locked yet. A read error other than a write cut short leaves the file for a
/// later pass. Failures are logged per file, so callers still read the sealed files.
pub(crate) fn recover_partial_feather_files(directory: &Path) {
    let suffix = format!(".{FEATHER_PARTIAL_EXTENSION}");
    let mut directories = vec![directory.to_path_buf()];

    while let Some(directory) = directories.pop() {
        let entries = match fs::read_dir(&directory) {
            Ok(entries) => entries,
            Err(e) if e.kind() == io::ErrorKind::NotFound => continue,
            Err(e) => {
                log::warn!(
                    "Failed to list {} for partial Feather files: {e}",
                    directory.display()
                );
                continue;
            }
        };

        for entry in entries.flatten() {
            let path = entry.path();

            if entry.file_type().is_ok_and(|file_type| file_type.is_dir()) {
                directories.push(path);
            } else if path.to_string_lossy().ends_with(&suffix)
                && let Err(e) = recover_partial_feather_file(&path)
            {
                log::warn!("Failed to recover Feather file {}: {e}", path.display());
            }
        }
    }
}

// Recovers a file its writer released after a failure, so its complete batches stay promotable
fn recover_released_feather_file(partial_path: &Path) {
    if let Err(e) = recover_partial_feather_file(partial_path) {
        log::error!(
            "Failed to recover Feather file {}: {e}",
            partial_path.display()
        );
    }
}

fn recover_partial_feather_file(partial_path: &Path) -> anyhow::Result<()> {
    let file = match OpenOptions::new().read(true).write(true).open(partial_path) {
        Ok(file) => file,
        Err(e) if e.kind() == io::ErrorKind::NotFound => return Ok(()),
        Err(e) => return Err(e.into()),
    };

    match file.try_lock() {
        Ok(()) => {}
        Err(TryLockError::WouldBlock) => return Ok(()),
        Err(TryLockError::Error(e)) => return Err(e.into()),
    }

    let len = file.metadata()?.len();

    // A writer renames its file before unlocking it, and a new file stays empty until its writer
    // locks it, so recovery leaves a missing or empty file alone
    if !partial_path.try_exists()? || len == 0 {
        return Ok(());
    }

    let mut batches = 0;
    let mut complete_len = 0;
    let mut tail_error = None;

    match StreamReader::try_new(BufReader::new(&file), None) {
        Ok(mut reader) => loop {
            match reader.next() {
                Some(Ok(_)) => {
                    batches += 1;
                    complete_len = reader.get_mut().stream_position()?;
                }
                Some(Err(e)) => {
                    tail_error = Some(e);
                    break;
                }
                None => break,
            }
        },
        Err(e) => tail_error = Some(e),
    }

    // Only a write cut short ends the stream inside a message; any other read error could hide
    // complete batches after it, so the file waits for a later pass
    if let Some(e) = tail_error.take_if(|e| {
        !matches!(e, ArrowError::IoError(_, source) if source.kind() == io::ErrorKind::UnexpectedEof)
    }) {
        return Err(e.into());
    }

    let cause = tail_error.map_or_else(String::new, |e| format!(": {e}"));

    if batches == 0 {
        fs::remove_file(partial_path)?;

        log::warn!(
            "Removed Feather file {} with no complete record batch{cause}",
            partial_path.display()
        );
        return Ok(());
    }

    let dropped = len.saturating_sub(complete_len);

    // Arrow C++ rejects trailing bytes that form no message, and EOF at a message boundary ends a
    // valid stream, so the file ends at its last complete batch
    if dropped > 0 {
        file.set_len(complete_len)?;
    }

    file.sync_all()?;

    let sealed_path = partial_path.with_extension("");
    fs::rename(partial_path, &sealed_path)?;

    let tail = if dropped == 0 {
        String::new()
    } else {
        format!(", dropped {dropped} trailing byte(s){cause}")
    };

    log::warn!(
        "Recovered {batches} record batch(es) into Feather file {}{tail}",
        sealed_path.display()
    );
    Ok(())
}

#[cfg(test)]
mod tests {
    use std::sync::{Arc, Mutex, atomic::Ordering};

    use nautilus_common::{clock::VirtualClock, live::LiveClock};
    #[cfg(target_os = "linux")]
    use nautilus_model::data::HasTsInit;
    use nautilus_model::{
        data::{Data, NautilusRecordType, QuoteTick, TradeTick},
        enums::AggressorSide,
        identifiers::{InstrumentId, TradeId},
        types::{ERROR_PRICE, Price, Quantity},
    };
    use nautilus_serialization::arrow::{
        ArrowSchemaProvider, DecodeDataFromRecordBatch, EncodeToRecordBatch,
    };
    use rstest::rstest;
    use tempfile::TempDir;

    use super::*;
    use crate::{
        common::datafusion::identifiers_from_record_batches,
        writer::materializer::restore_staged_record_batches,
    };

    fn feather_files(directory: &Path, extension: &str) -> Vec<PathBuf> {
        let mut files = Vec::new();
        let mut stack = vec![directory.to_path_buf()];

        while let Some(path) = stack.pop() {
            for entry in fs::read_dir(&path).unwrap().flatten() {
                let path = entry.path();
                if path.is_dir() {
                    stack.push(path);
                } else if path.to_string_lossy().ends_with(&format!(".{extension}")) {
                    files.push(path);
                }
            }
        }

        files.sort();
        files
    }

    // Reads a type's file and splits it into batches that each carry their rows' metadata
    fn read_feather_batches(path: &Path) -> Vec<RecordBatch> {
        StreamReader::try_new(File::open(path).unwrap(), None)
            .unwrap()
            .map(Result::unwrap)
            .flat_map(|batch| restore_staged_record_batches(batch).unwrap())
            .collect()
    }

    fn read_quotes(path: &Path) -> Vec<Data> {
        read_feather_batches(path)
            .into_iter()
            .flat_map(|batch| {
                let metadata = batch.schema().metadata().clone();
                QuoteTick::decode_data_batch(&metadata, batch).unwrap()
            })
            .collect()
    }

    fn quote_at(ts: u64) -> QuoteTick {
        QuoteTick::new(
            InstrumentId::from("AUD/USD.SIM"),
            Price::from("1.0"),
            Price::from("1.1"),
            Quantity::from("1000"),
            Quantity::from("1000"),
            UnixNanos::from(ts),
            UnixNanos::from(ts),
        )
    }

    fn quotes_at(timestamps: &[u64]) -> Vec<Data> {
        timestamps
            .iter()
            .map(|&ts| Data::from(quote_at(ts)))
            .collect()
    }

    // Seals one flushed batch per quote, then reopens the stream as a partial file whose writer
    // exited before sealing it
    fn write_partial_file(directory: &Path, timestamps: &[u64]) -> PathBuf {
        let mut writer = FeatherWriter::new(
            directory.to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            Some(0),
        );

        for &ts in timestamps {
            writer.write(quote_at(ts)).unwrap();
            writer.flush().unwrap();
        }

        writer.close().unwrap();

        let sealed = feather_files(directory, FEATHER_EXTENSION).remove(0);
        let partial = sealed.with_extension(FEATHER_PARTIAL_EXTENSION);
        fs::rename(&sealed, &partial).unwrap();
        partial
    }

    const TORN_TAIL: &str = ": Io error: failed to fill whole buffer";

    static RECOVERY_LOGS: RecoveryLogCapture = RecoveryLogCapture(Mutex::new(Vec::new()));

    // The logger is process-global, so only warnings naming `directory` belong to the caller
    fn recover_with_warnings(directory: &Path) -> Vec<String> {
        let _ = log::set_logger(&RECOVERY_LOGS);
        log::set_max_level(log::LevelFilter::Warn);

        recover_partial_feather_files(directory);

        let directory = directory.display().to_string();
        RECOVERY_LOGS
            .0
            .lock()
            .unwrap()
            .iter()
            .filter(|(level, message)| *level == log::Level::Warn && message.contains(&directory))
            .map(|(_, message)| message.clone())
            .collect()
    }

    struct RecoveryLogCapture(Mutex<Vec<(log::Level, String)>>);

    impl log::Log for RecoveryLogCapture {
        fn enabled(&self, metadata: &log::Metadata<'_>) -> bool {
            metadata.target() == "nautilus_persistence::writer::feather"
        }

        fn log(&self, record: &log::Record<'_>) {
            if self.enabled(record.metadata()) {
                self.0
                    .lock()
                    .unwrap()
                    .push((record.level(), record.args().to_string()));
            }
        }

        fn flush(&self) {}
    }

    #[rstest]
    fn test_subscription_receives_typed_quotes_and_unsubscribes() {
        use nautilus_common::msgbus::{MStr, publish_quote};

        let temp_dir = TempDir::new().unwrap();

        let writer = Rc::new(RefCell::new(FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            Some(0),
        )));

        let quote = QuoteTick::new(
            InstrumentId::from("AUD/USD.SIM"),
            Price::from("1.0"),
            Price::from("1.1"),
            Quantity::from("2"),
            Quantity::from("3"),
            4.into(),
            5.into(),
        );
        let handler = FeatherWriter::subscribe_to_message_bus(writer.clone()).unwrap();
        publish_quote(MStr::topic("data.quotes.AUD/USD.SIM").unwrap(), &quote);
        FeatherWriter::unsubscribe_from_message_bus(&handler);
        publish_quote(MStr::topic("data.quotes.AUD/USD.SIM").unwrap(), &quote);
        writer.borrow_mut().close().unwrap();

        let rows = feather_files(temp_dir.path(), FEATHER_EXTENSION)
            .iter()
            .flat_map(|path| read_feather_batches(path))
            .map(|batch| batch.num_rows())
            .sum::<usize>();
        assert_eq!(rows, 1);
    }

    #[rstest]
    fn test_subscription_records_greeks_once_and_unsubscribes() {
        use nautilus_common::msgbus::{MStr, publish_any, publish_greeks};
        use nautilus_model::data::greeks::GreeksData;
        use nautilus_serialization::arrow::DecodeTypedFromRecordBatch;
        let temp_dir = TempDir::new().unwrap();

        let writer = Rc::new(RefCell::new(FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            Some(0),
        )));
        let greeks = GreeksData::from_delta("AUD/USD.SIM".parse().unwrap(), 1.25, 8.0, 11.into());
        let portfolio = greeks.to_portfolio_greeks();
        let greeks_topic = MStr::topic("data.GreeksData.instrument_id=AUD/USD.SIM").unwrap();
        let portfolio_topic = MStr::topic("data.PortfolioGreeks").unwrap();
        let handler = FeatherWriter::subscribe_to_message_bus(writer.clone()).unwrap();
        publish_greeks(greeks_topic, &greeks);
        publish_any(portfolio_topic, &portfolio);
        FeatherWriter::unsubscribe_from_message_bus(&handler);
        publish_greeks(greeks_topic, &greeks);
        publish_any(portfolio_topic, &portfolio);
        writer.borrow_mut().close().unwrap();
        let mut counts = HashMap::new();

        for batch in feather_files(temp_dir.path(), FEATHER_EXTENSION)
            .iter()
            .flat_map(|path| read_feather_batches(path))
        {
            let schema = batch.schema();
            let name = schema.metadata()["type_name"].clone();
            *counts.entry(name.clone()).or_insert(0) += batch.num_rows();
            match name.as_str() {
                "GreeksData" => {
                    let values = GreeksData::decode_typed_batch(schema.metadata(), batch).unwrap();
                    assert_eq!(values[0].instrument_id, greeks.instrument_id);
                    assert_eq!(values[0].greeks, greeks.greeks);
                    assert_eq!(values[0].ts_init, greeks.ts_init);
                }
                "PortfolioGreeks" => {
                    let values =
                        PortfolioGreeks::decode_typed_batch(schema.metadata(), batch).unwrap();
                    assert_eq!(values[0].greeks, portfolio.greeks);
                    assert_eq!(values[0].ts_init, portfolio.ts_init);
                }
                _ => panic!("unexpected record type {name}"),
            }
        }

        assert_eq!(
            counts,
            HashMap::from([
                ("GreeksData".to_string(), 1),
                ("PortfolioGreeks".to_string(), 1),
            ])
        );
    }

    #[rstest]
    #[case(false)]
    #[case(true)]
    fn test_message_write_error_reaches_flush_or_close(#[case] close: bool) {
        let temp_dir = TempDir::new().unwrap();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            Some(0),
        );

        let quote = QuoteTick::new(
            InstrumentId::from("AUD/USD.SIM"),
            ERROR_PRICE,
            ERROR_PRICE,
            Quantity::from("1"),
            Quantity::from("2"),
            3.into(),
            4.into(),
        );
        let write_error = writer.write_any_message(&quote).unwrap_err().to_string();

        let error = if close {
            StreamingSink::close(&mut writer)
        } else {
            StreamingSink::flush(&mut writer)
        }
        .unwrap_err();

        assert_eq!(error.to_string(), write_error);
    }

    #[rstest]
    fn test_writer_manager_keys() {
        let temp_dir = TempDir::new().unwrap();

        // Create a test time source
        let clock = WriterClock::Test(Arc::new(AtomicU64::new(0)));
        let timestamp = clock.timestamp_ns();

        let mut manager = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            None, // flush_interval_ms
        );

        let instrument_id = "AAPL.AAPL";

        // Write a dummy value
        let quote = QuoteTick::new(
            InstrumentId::from(instrument_id),
            Price::from("100.0"),
            Price::from("100.0"),
            Quantity::from("100.0"),
            Quantity::from("100.0"),
            UnixNanos::from(1_000_000_000_000_000_000),
            UnixNanos::from(1_000_000_000_000_000_000),
        );

        let trade = TradeTick::new(
            InstrumentId::from(instrument_id),
            Price::from("100.0"),
            Quantity::from("100.0"),
            AggressorSide::Buy,
            TradeId::from("1"),
            UnixNanos::from(1_000_000_000_000_000_000),
            UnixNanos::from(1_000_000_000_000_000_000),
        );

        manager.write(quote).unwrap();
        manager.write(trade).unwrap();

        // Check keys and paths for quotes and trades
        let path = manager.get_writer_path(QuoteTick::catalog_family());
        let expected_path = temp_dir
            .path()
            .join("quotes")
            .join(format!("quotes_{timestamp}.feather"));
        assert_eq!(path.path, expected_path);
        assert!(manager.writers.contains_key(&path));
        let writer = manager.writers.get(&path).unwrap();
        assert!(writer.size() > 0);

        let path = manager.get_writer_path(TradeTick::catalog_family());
        let expected_path = temp_dir
            .path()
            .join("trades")
            .join(format!("trades_{timestamp}.feather"));
        assert_eq!(path.path, expected_path);
        assert!(manager.writers.contains_key(&path));
        let writer = manager.writers.get(&path).unwrap();
        assert!(writer.size() > 0);
    }

    #[rstest]
    #[case::quotes(NautilusDataType::QuoteTick.into(), "quotes/quotes_0.feather")]
    #[case::record(NautilusRecordType::OrderFilled.into(), "order_filled/order_filled_0.feather")]
    #[case::instrument_class(
        CatalogDataType::Instrument(NautilusInstrumentType::CurrencyPair),
        "instruments/currency_pair/instruments_0.feather"
    )]
    #[case::custom(
        NautilusDataType::Custom { type_name: "Signal".to_string() }.into(),
        "data/custom/Signal/Signal_0.feather"
    )]
    fn writer_paths_hold_one_file_per_type(
        #[case] data_type: CatalogDataType,
        #[case] expected: &str,
    ) {
        let temp_dir = TempDir::new().unwrap();

        let manager = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            None,
        );

        let path = manager.build_writer_path(&data_type, UnixNanos::default(), 0);

        assert_eq!(path, temp_dir.path().join(expected));
    }

    #[rstest]
    fn existing_feather_writer_implements_streaming_data_sink() {
        let temp_dir = TempDir::new().unwrap();
        let clock = WriterClock::Test(Arc::new(AtomicU64::new(0)));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            None,
        );

        let quote = QuoteTick::new(
            InstrumentId::from("AUD/USD.SIM"),
            Price::from("1.0"),
            Price::from("1.1"),
            Quantity::from("1000"),
            Quantity::from("1000"),
            UnixNanos::from(1_000),
            UnixNanos::from(1_000),
        );

        StreamingSink::write_data(&mut writer, Data::Quote(quote)).unwrap();
        StreamingSink::flush(&mut writer).unwrap();
        StreamingSink::close(&mut writer).unwrap();

        let files = feather_files(&temp_dir.path().join("quotes"), FEATHER_EXTENSION);
        assert_eq!(files.len(), 1);
        let identifiers =
            identifiers_from_record_batches(&read_feather_batches(&files[0])).unwrap();
        assert_eq!(identifiers, vec!["AUD/USD.SIM".to_string()]);
    }

    #[rstest]
    fn scheduled_rotation_keeps_time_of_day_anchor_after_late_rotation() {
        let temp_dir = TempDir::new().unwrap();

        let now = Arc::new(AtomicU64::new(
            1_767_258_000_000_000_000, // 2026-01-01 09:00:00 UTC
        ));

        let path = FileWriterPath {
            path: PathBuf::from("quotes.feather"),
            data_type: NautilusDataType::QuoteTick.into(),
        };

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::clone(&now)),
            RotationConfig::ScheduledDates {
                interval_ns: 86_400_000_000_000,
                rotation_time: UnixNanos::from(36_000_000_000_000u64),
                rotation_timezone: jiff::tz::TimeZone::UTC,
            },
            None,
            None,
        );

        assert!(!writer.check_scheduled_rotation(&path));
        assert_eq!(
            writer.next_rotation_times[&path],
            UnixNanos::from(1_767_261_600_000_000_000u64),
        );

        now.store(1_767_263_400_000_000_000, Ordering::Relaxed); // 10:30 UTC
        assert!(writer.check_scheduled_rotation(&path));
        assert_eq!(
            writer.next_rotation_times[&path],
            UnixNanos::from(1_767_348_000_000_000_000u64),
        );
    }

    #[rstest]
    fn test_file_writer_round_trip() {
        let instrument_id = "AAPL.AAPL";

        // Write a dummy value.
        let quote = QuoteTick::new(
            InstrumentId::from(instrument_id),
            Price::from("100.0"),
            Price::from("100.0"),
            Quantity::from("100.0"),
            Quantity::from("100.0"),
            UnixNanos::from(100),
            UnixNanos::from(100),
        );
        let metadata = QuoteTick::metadata(&quote);
        let schema = QuoteTick::get_schema(Some(metadata.clone()));
        let batch = QuoteTick::encode_batch(&QuoteTick::metadata(&quote), &[quote]).unwrap();

        let temp_dir = TempDir::new().unwrap();
        let path = temp_dir.path().join("quotes.feather");
        let mut file = FeatherFile::create(&path, &schema, &RotationConfig::NoRotation).unwrap();
        file.write_record_batch(&batch).unwrap();
        file.seal().unwrap();

        assert!(!path.with_extension(FEATHER_PARTIAL_EXTENSION).exists());
        let mut reader = StreamReader::try_new(File::open(&path).unwrap(), None).unwrap();

        let read_metadata = reader.schema().metadata().clone();
        let mut expected_metadata = metadata.clone();
        expected_metadata.insert("type_name".to_string(), "QuoteTick".to_string());
        assert_eq!(read_metadata, expected_metadata);

        let read_batch = reader.next().unwrap().unwrap();
        assert_eq!(read_batch.column(0), batch.column(0));

        let decoded = QuoteTick::decode_data_batch(&metadata, batch).unwrap();
        assert_eq!(decoded[0], Data::from(quote));
    }

    #[rstest]
    fn test_round_trip() {
        // Create a temporary directory for base path
        let temp_dir = TempDir::new_in(".").unwrap();

        // Create a test time source
        let clock = WriterClock::Test(Arc::new(AtomicU64::new(0)));

        let mut manager = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            None, // flush_interval_ms
        );

        let instrument_id = "AAPL.AAPL";

        // Write a dummy value.
        let quote = QuoteTick::new(
            InstrumentId::from(instrument_id),
            Price::from("100.0"),
            Price::from("100.0"),
            Quantity::from("100.0"),
            Quantity::from("100.0"),
            UnixNanos::from(100),
            UnixNanos::from(100),
        );

        let trade = TradeTick::new(
            InstrumentId::from(instrument_id),
            Price::from("100.0"),
            Quantity::from("100.0"),
            AggressorSide::Buy,
            TradeId::from("1"),
            UnixNanos::from(100),
            UnixNanos::from(100),
        );

        manager.write(quote).unwrap();
        manager.write(trade).unwrap();

        let paths = manager.writers.keys().cloned().collect::<Vec<_>>();
        assert_eq!(paths.len(), 2);

        manager.close().unwrap();

        let mut recovered_quotes = Vec::new();
        let mut recovered_trades = Vec::new();

        for path in paths {
            let path_str = path.path;
            for batch in read_feather_batches(&path_str) {
                let metadata = batch.schema().metadata().clone();

                if path_str.to_str().unwrap().contains("quotes") {
                    let decoded = QuoteTick::decode_data_batch(&metadata, batch).unwrap();
                    recovered_quotes.extend(decoded);
                } else if path_str.to_str().unwrap().contains("trades") {
                    let decoded = TradeTick::decode_data_batch(&metadata, batch).unwrap();
                    recovered_trades.extend(decoded);
                }
            }
        }

        // Assert that the recovered data matches the written data
        assert_eq!(recovered_quotes.len(), 1, "Expected one QuoteTick record");
        assert_eq!(recovered_trades.len(), 1, "Expected one TradeTick record");

        // Check key fields to ensure the data round-tripped correctly
        assert_eq!(recovered_quotes[0], Data::from(quote));
        assert_eq!(recovered_trades[0], Data::from(trade));
    }

    #[rstest]
    fn test_write_data_enum() {
        let temp_dir = TempDir::new().unwrap();
        let clock = WriterClock::Test(Arc::new(AtomicU64::new(0)));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            None,
        );

        let quote = QuoteTick::new(
            InstrumentId::from("AUD/USD.SIM"),
            Price::from("1.0"),
            Price::from("1.0"),
            Quantity::from("1000"),
            Quantity::from("1000"),
            UnixNanos::from(1000),
            UnixNanos::from(1000),
        );

        // Test writing via write_data
        writer.write_data(Data::Quote(quote)).unwrap();
        writer.flush().unwrap();

        // Verify file was created
        assert!(!writer.writers.is_empty() || temp_dir.path().read_dir().unwrap().count() > 0);
    }

    #[rstest]
    fn test_write_data_all_types() {
        let temp_dir = TempDir::new().unwrap();
        let clock = WriterClock::Test(Arc::new(AtomicU64::new(0)));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            None,
        );

        let instrument_id = InstrumentId::from("AUD/USD.SIM");

        // Test all data types
        let quote = QuoteTick::new(
            instrument_id,
            Price::from("1.0"),
            Price::from("1.0"),
            Quantity::from("1000"),
            Quantity::from("1000"),
            UnixNanos::from(1000),
            UnixNanos::from(1000),
        );
        writer.write_data(Data::Quote(quote)).unwrap();

        let trade = TradeTick::new(
            instrument_id,
            Price::from("1.0"),
            Quantity::from("1000"),
            AggressorSide::Buy,
            TradeId::from("1"),
            UnixNanos::from(2000),
            UnixNanos::from(2000),
        );
        writer.write_data(Data::Trade(trade)).unwrap();

        let delta = OrderBookDelta::clear(
            instrument_id,
            0,
            UnixNanos::from(3000),
            UnixNanos::from(3000),
        );
        writer.write_data(Data::BookDelta(delta)).unwrap();

        writer.flush().unwrap();
    }

    #[rstest]
    fn test_auto_flush() {
        let temp_dir = TempDir::new().unwrap();
        let shared_time = Arc::new(AtomicU64::new(0));
        let clock = WriterClock::Test(Arc::clone(&shared_time));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            Some(100), // 100ms flush interval
        );

        let quote = QuoteTick::new(
            InstrumentId::from("AUD/USD.SIM"),
            Price::from("1.0"),
            Price::from("1.0"),
            Quantity::from("1000"),
            Quantity::from("1000"),
            UnixNanos::from(1000),
            UnixNanos::from(1000),
        );

        // Write first quote; interval not elapsed, so nothing is flushed yet
        writer.write(quote).unwrap();
        assert_eq!(writer.writers.len(), 1);
        assert_eq!(writer.last_flush_ns, UnixNanos::from(0));

        // Advance the shared time source past the 100ms flush interval
        shared_time.store(200_000_000, Ordering::Relaxed);

        // Second write hits the flush boundary: both quotes reach the same open file
        let quote2 = QuoteTick::new(
            InstrumentId::from("AUD/USD.SIM"),
            Price::from("1.1"),
            Price::from("1.1"),
            Quantity::from("1000"),
            Quantity::from("1000"),
            UnixNanos::from(2000),
            UnixNanos::from(2000),
        );
        writer.write(quote2).unwrap();

        // Windows locks block reads through other handles, so the open file is measured instead:
        // if closing adds only the 8-byte end-of-stream marker, both quotes were already on disk
        let partial = feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION);
        let flushed_len = fs::metadata(&partial[0]).unwrap().len();
        let open_writers = writer.writers.len();
        let last_flush_ns = writer.last_flush_ns;
        let sealed_before_close = feather_files(temp_dir.path(), FEATHER_EXTENSION);
        writer.close().unwrap();

        let sealed = partial[0].with_extension("");
        let sealed_rows = read_feather_batches(&sealed)
            .iter()
            .map(RecordBatch::num_rows)
            .sum::<usize>();
        assert_eq!(open_writers, 1);
        assert_eq!(last_flush_ns, UnixNanos::from(200_000_000));
        assert_eq!(partial.len(), 1);
        assert!(sealed_before_close.is_empty());
        assert_eq!(fs::metadata(&sealed).unwrap().len(), flushed_len + 8);
        assert_eq!(sealed_rows, 2);
    }

    #[rstest]
    fn flushes_append_to_one_file_per_type_until_close() {
        let temp_dir = TempDir::new().unwrap();
        let shared_time = Arc::new(AtomicU64::new(0));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::clone(&shared_time)),
            RotationConfig::NoRotation,
            None,
            Some(1_000),
        );

        let quotes = (1..=10)
            .map(|minute| {
                QuoteTick::new(
                    InstrumentId::from("AUD/USD.SIM"),
                    Price::from("1.0"),
                    Price::from("1.1"),
                    Quantity::from("1000"),
                    Quantity::from("1000"),
                    UnixNanos::from(minute * 60_000_000_000),
                    UnixNanos::from(minute * 60_000_000_000),
                )
            })
            .collect::<Vec<_>>();

        // Each write crosses a flush interval, as a 1-minute bar does in a backtest
        for quote in &quotes {
            shared_time.store(quote.ts_init.as_u64(), Ordering::Relaxed);
            writer.write(*quote).unwrap();
        }

        let partial_before_close = feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION);
        let sealed_before_close = feather_files(temp_dir.path(), FEATHER_EXTENSION);
        writer.close().unwrap();
        let sealed = feather_files(temp_dir.path(), FEATHER_EXTENSION);

        let recovered = read_feather_batches(&sealed[0])
            .into_iter()
            .flat_map(|batch| {
                let metadata = batch.schema().metadata().clone();
                QuoteTick::decode_data_batch(&metadata, batch).unwrap()
            })
            .collect::<Vec<_>>();

        assert_eq!(partial_before_close.len(), 1);
        assert!(sealed_before_close.is_empty());
        assert_eq!(sealed.len(), 1);
        assert!(feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION).is_empty());
        assert_eq!(
            recovered,
            quotes.into_iter().map(Data::from).collect::<Vec<_>>()
        );
    }

    #[rstest]
    #[case::below_row_cap(3, vec![3])]
    #[case::above_row_cap(
        DEFAULT_DATA_BATCH_CHUNK_SIZE + 1,
        vec![DEFAULT_DATA_BATCH_CHUNK_SIZE, 1]
    )]
    fn writes_between_flushes_append_one_batch_per_row_cap(
        #[case] count: usize,
        #[case] expected_batch_rows: Vec<usize>,
    ) {
        let temp_dir = TempDir::new().unwrap();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            Some(0),
        );
        let instrument_ids = [
            InstrumentId::from("AUD/USD.SIM"),
            InstrumentId::from("EUR/USD.SIM"),
        ];

        for index in 0..count {
            writer
                .write(QuoteTick::new(
                    instrument_ids[index % 2],
                    Price::from("1.0"),
                    Price::from("1.1"),
                    Quantity::from("1000"),
                    Quantity::from("1000"),
                    UnixNanos::from(index as u64),
                    UnixNanos::from(index as u64),
                ))
                .unwrap();
        }

        writer.close().unwrap();
        let sealed = feather_files(temp_dir.path(), FEATHER_EXTENSION);
        let batch_rows = StreamReader::try_new(File::open(&sealed[0]).unwrap(), None)
            .unwrap()
            .map(|batch| batch.unwrap().num_rows())
            .collect::<Vec<_>>();

        assert_eq!(sealed.len(), 1);
        assert_eq!(batch_rows, expected_batch_rows);
    }

    #[rstest]
    fn drop_seals_open_files() {
        let temp_dir = TempDir::new().unwrap();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            None,
        );
        writer
            .write(QuoteTick::new(
                InstrumentId::from("AUD/USD.SIM"),
                Price::from("1.0"),
                Price::from("1.1"),
                Quantity::from("1000"),
                Quantity::from("1000"),
                UnixNanos::from(1_000),
                UnixNanos::from(1_000),
            ))
            .unwrap();

        drop(writer);

        assert_eq!(feather_files(temp_dir.path(), FEATHER_EXTENSION).len(), 1);
        assert!(feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION).is_empty());
    }

    #[rstest]
    fn reserved_paths_skip_files_left_in_the_directory() {
        let temp_dir = TempDir::new().unwrap();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(5))),
            RotationConfig::NoRotation,
            None,
            None,
        );
        let sealed = temp_dir.path().join("trades").join("trades_5.feather");
        let partial = temp_dir
            .path()
            .join("trades")
            .join("trades_5-1.feather.partial");
        fs::create_dir_all(sealed.parent().unwrap()).unwrap();
        fs::write(&sealed, b"sealed").unwrap();
        fs::write(&partial, b"partial").unwrap();

        let path = writer.reserve_writer_path(NautilusDataType::TradeTick.into());

        assert_eq!(
            path.path,
            temp_dir.path().join("trades").join("trades_5-2.feather")
        );
        assert_eq!(fs::read(&sealed).unwrap(), b"sealed");
        assert_eq!(fs::read(&partial).unwrap(), b"partial");
    }

    #[rstest]
    #[case::same_time(vec![5; 32])]
    #[case::clock_rewinds(vec![5, 3, 5])]
    #[case::clock_advances_then_rewinds(vec![5, 7, 5, 7])]
    fn path_reservations_are_bounded_after_files_are_deleted(#[case] timestamps: Vec<u64>) {
        let temp_dir = TempDir::new().unwrap();
        let time = Arc::new(AtomicU64::new(0));
        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::clone(&time)),
            RotationConfig::NoRotation,
            None,
            None,
        );
        let mut paths = HashSet::new();

        for timestamp in &timestamps {
            time.store(*timestamp, Ordering::Relaxed);
            writer.write(quote_at(1)).unwrap();
            writer.close().unwrap();
            let path = feather_files(temp_dir.path(), FEATHER_EXTENSION).remove(0);
            assert!(paths.insert(path.clone()));
            fs::remove_file(path).unwrap();
        }

        assert_eq!(paths.len(), timestamps.len());
        assert_eq!(writer.path_sequences.len(), 1);
    }

    #[rstest]
    fn test_close() {
        let temp_dir = TempDir::new().unwrap();
        let clock = WriterClock::Test(Arc::new(AtomicU64::new(0)));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            None,
        );

        let quote = QuoteTick::new(
            InstrumentId::from("AUD/USD.SIM"),
            Price::from("1.0"),
            Price::from("1.0"),
            Quantity::from("1000"),
            Quantity::from("1000"),
            UnixNanos::from(1000),
            UnixNanos::from(1000),
        );

        writer.write(quote).unwrap();
        assert!(!writer.writers.is_empty());

        writer.close().unwrap();
        assert!(writer.writers.is_empty());
    }

    #[rstest]
    fn test_write_data_orderbook_deltas() {
        let temp_dir = TempDir::new().unwrap();
        let clock = WriterClock::Test(Arc::new(AtomicU64::new(0)));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            None,
        );

        let instrument_id = InstrumentId::from("AUD/USD.SIM");
        let delta1 = OrderBookDelta::clear(
            instrument_id,
            0,
            UnixNanos::from(1000),
            UnixNanos::from(1000),
        );
        let delta2 = OrderBookDelta::clear(
            instrument_id,
            0,
            UnixNanos::from(2000),
            UnixNanos::from(2000),
        );

        let deltas = OrderBookDeltas::new(instrument_id, vec![delta1, delta2]);
        // Test writing OrderBookDeltas via write_data
        writer
            .write_data(Data::BookDeltas(Box::new(deltas)))
            .unwrap();
        writer.flush().unwrap();
    }

    #[rstest]
    fn feather_writer_is_send() {
        fn assert_send<T: Send>() {}
        assert_send::<FeatherWriter>();
        assert_send::<WriterClock>();
    }

    #[rstest]
    fn writer_clock_test_source_reads_shared_atomic() {
        let shared = Arc::new(AtomicU64::new(7));
        let clock = WriterClock::Test(Arc::clone(&shared));
        assert_eq!(clock.timestamp_ns(), UnixNanos::from(7));

        shared.store(42, Ordering::Relaxed);
        assert_eq!(clock.timestamp_ns(), UnixNanos::from(42));
    }

    #[rstest]
    fn writer_clock_from_shared_clock_wires_test_clock() {
        let test_clock = Rc::new(RefCell::new(VirtualClock::new()));
        test_clock
            .borrow_mut()
            .advance_time(UnixNanos::from(42), true);
        let clock: Rc<RefCell<dyn Clock>> = test_clock.clone();

        let (writer_clock, shared) = WriterClock::from_shared_clock(&clock);
        let shared = shared.expect("non-live clocks must return a shared atomic");

        // Seeded with the source clock's current time
        assert_eq!(writer_clock.timestamp_ns(), UnixNanos::from(42));

        // Bridge refresh: advancing the source clock and storing into the atomic
        // is what the PyO3 wrapper does before each forwarded call
        test_clock
            .borrow_mut()
            .advance_time(UnixNanos::from(99), true);
        shared.store(clock.borrow().timestamp_ns().as_u64(), Ordering::Relaxed);
        assert_eq!(writer_clock.timestamp_ns(), UnixNanos::from(99));
    }

    #[rstest]
    fn writer_clock_from_shared_clock_live_clock_is_live() {
        let clock: Rc<RefCell<dyn Clock>> = Rc::new(RefCell::new(LiveClock::new(None)));

        let (writer_clock, shared) = WriterClock::from_shared_clock(&clock);

        assert!(matches!(writer_clock, WriterClock::Live));
        assert!(shared.is_none());
    }

    #[rstest]
    #[cfg(unix)]
    fn failed_seal_of_a_removed_file_reports_the_error_and_publishes_nothing() {
        let temp_dir = TempDir::new().unwrap();
        let (sealed_tx, sealed_rx) = std::sync::mpsc::channel();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            None,
        )
        .with_sealed_sender(sealed_tx);
        writer
            .write(QuoteTick::new(
                InstrumentId::from("AUD/USD.SIM"),
                Price::from("1.0"),
                Price::from("1.1"),
                Quantity::from("1000"),
                Quantity::from("1000"),
                UnixNanos::from(1_000),
                UnixNanos::from(1_000),
            ))
            .unwrap();

        // Removing the open partial file makes the seal's rename fail after the stream finishes
        let partial = feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION);
        fs::remove_file(&partial[0]).unwrap();

        let error = writer.close().unwrap_err();

        assert_eq!(error.to_string(), "No such file or directory (os error 2)");
        assert!(writer.is_closed());
        assert!(feather_files(temp_dir.path(), FEATHER_EXTENSION).is_empty());
        assert_eq!(
            sealed_rx.try_iter().collect::<Vec<_>>(),
            Vec::<PathBuf>::new()
        );
    }

    #[rstest]
    #[cfg(target_os = "linux")]
    fn failed_seal_recovers_the_batches_that_reached_the_file() {
        let temp_dir = TempDir::new().unwrap();
        let (sealed_tx, sealed_rx) = std::sync::mpsc::channel();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            Some(0),
        )
        .with_sealed_sender(sealed_tx);

        writer.write(quote_at(1)).unwrap();
        writer.flush().unwrap();
        writer.write(quote_at(2)).unwrap();

        // Route the open stream to a device that reports a full disk, so sealing it fails
        let (path, file) = writer.writers.iter_mut().next().unwrap();
        let sealed = path.path.clone();
        let full = OpenOptions::new().write(true).open("/dev/full").unwrap();
        file.writer =
            StreamWriter::try_new(CountingWriter::new(BufWriter::new(full)), &file.schema).unwrap();

        let error = writer.close().unwrap_err();

        assert_eq!(
            error.to_string(),
            "Io error: No space left on device (os error 28)"
        );
        assert!(writer.is_closed());
        assert!(feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION).is_empty());
        assert_eq!(read_quotes(&sealed), quotes_at(&[1]));
        assert_eq!(sealed_rx.try_iter().collect::<Vec<_>>(), vec![sealed]);
    }

    #[rstest]
    #[case::nothing_flushed(false, vec![3])]
    #[case::batch_flushed(true, vec![1, 3])]
    #[cfg(target_os = "linux")]
    fn failed_flush_recovers_the_abandoned_file_and_next_write_opens_a_new_one(
        #[case] flushed: bool,
        #[case] expected: Vec<u64>,
    ) {
        let temp_dir = TempDir::new().unwrap();
        let (sealed_tx, sealed_rx) = std::sync::mpsc::channel();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            Some(0),
        )
        .with_sealed_sender(sealed_tx);

        writer.write(quote_at(1)).unwrap();

        if flushed {
            writer.flush().unwrap();
        }

        writer.write(quote_at(2)).unwrap();

        // Route the open stream to a device that reports a full disk on every write
        let (abandoned, file) = writer.writers.iter_mut().next().unwrap();
        let abandoned = abandoned.clone();
        let full = OpenOptions::new().write(true).open("/dev/full").unwrap();
        file.writer =
            StreamWriter::try_new(CountingWriter::new(BufWriter::new(full)), &file.schema).unwrap();

        let error = writer.flush().unwrap_err();
        writer.write(quote_at(3)).unwrap();
        writer.close().unwrap();

        let mut recovered = feather_files(temp_dir.path(), FEATHER_EXTENSION)
            .iter()
            .flat_map(|path| read_quotes(path))
            .collect::<Vec<_>>();
        recovered.sort_by_key(HasTsInit::ts_init);

        // Quote 2 was buffered when the flush failed, and the full device also fails the writer's
        // drop-time flush, so it reaches no file
        assert_eq!(
            error.to_string(),
            "Io error: No space left on device (os error 28)"
        );
        assert_eq!(abandoned.path.exists(), flushed);
        assert!(feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION).is_empty());
        assert_eq!(recovered, quotes_at(&expected));
        let mut notified = sealed_rx.try_iter().collect::<Vec<_>>();
        notified.sort();
        assert_eq!(notified, feather_files(temp_dir.path(), FEATHER_EXTENSION));
    }

    #[rstest]
    #[case::end_of_stream_marker(0, &[], vec![1, 2, 3], "")]
    #[case::no_end_of_stream_marker(8, &[], vec![1, 2, 3], "")]
    #[case::stray_bytes(8, &[0xFF, 0xFF], vec![1, 2, 3], "")]
    #[case::truncated_last_batch(9, &[], vec![1, 2], TORN_TAIL)]
    fn recovery_seals_complete_batches_and_drops_the_rest(
        #[case] cut: usize,
        #[case] stray: &[u8],
        #[case] expected: Vec<u64>,
        #[case] cause: &str,
    ) {
        let temp_dir = TempDir::new().unwrap();
        let directory = temp_dir.path().join("recover");
        let partial = write_partial_file(&directory, &[1, 2, 3]);
        let mut staged = fs::read(&partial).unwrap();
        staged.truncate(staged.len() - cut);
        staged.extend_from_slice(stray);
        fs::write(&partial, &staged).unwrap();

        // The expected batches alone form this prefix, which ends before the end-of-stream marker
        let reference = fs::read(write_partial_file(
            &temp_dir.path().join("reference"),
            &expected,
        ))
        .unwrap();
        let complete = &reference[..reference.len() - 8];

        let warnings = recover_with_warnings(&directory);

        let sealed = partial.with_extension("");
        let dropped = staged.len() - complete.len();

        let tail = if dropped == 0 {
            String::new()
        } else {
            format!(", dropped {dropped} trailing byte(s){cause}")
        };

        assert!(!partial.exists());
        assert_eq!(fs::read(&sealed).unwrap(), complete);
        assert_eq!(read_quotes(&sealed), quotes_at(&expected));
        assert_eq!(
            warnings,
            vec![format!(
                "Recovered {} record batch(es) into Feather file {}{tail}",
                expected.len(),
                sealed.display()
            )],
        );
    }

    #[rstest]
    #[case::schema_only(0, "")]
    #[case::truncated_schema(1, TORN_TAIL)]
    fn recovery_removes_a_file_without_a_complete_batch(#[case] cut: u64, #[case] cause: &str) {
        let temp_dir = TempDir::new().unwrap();
        let partial = temp_dir.path().join("quotes_1.feather.partial");
        let file = File::create(&partial).unwrap();
        StreamWriter::try_new(&file, &QuoteTick::get_schema(None)).unwrap();
        file.set_len(file.metadata().unwrap().len() - cut).unwrap();
        drop(file);

        let warnings = recover_with_warnings(temp_dir.path());

        assert_eq!(fs::read_dir(temp_dir.path()).unwrap().count(), 0);
        assert_eq!(
            warnings,
            vec![format!(
                "Removed Feather file {} with no complete record batch{cause}",
                partial.display()
            )],
        );
    }

    #[rstest]
    fn recovery_leaves_a_file_with_an_unreadable_tail() {
        let temp_dir = TempDir::new().unwrap();
        let partial = write_partial_file(temp_dir.path(), &[1, 2]);

        // A continuation marker followed by a negative metadata length is corrupt, not cut short
        let mut staged = fs::read(&partial).unwrap();
        staged.truncate(staged.len() - 8);
        staged.extend_from_slice(&[0xFF, 0xFF, 0xFF, 0xFF, 0xF0, 0xFF, 0xFF, 0xFF]);
        fs::write(&partial, &staged).unwrap();

        let warnings = recover_with_warnings(temp_dir.path());

        assert_eq!(fs::read(&partial).unwrap(), staged);
        assert_eq!(
            warnings,
            vec![format!(
                "Failed to recover Feather file {}: Parser error: Invalid metadata length: -16",
                partial.display()
            )],
        );
    }

    #[rstest]
    #[cfg(unix)]
    fn recovery_continues_after_a_file_it_cannot_recover() {
        let temp_dir = TempDir::new().unwrap();
        let run = temp_dir.path().join("run");
        let partial = write_partial_file(&run, &[1]);

        // A partial path that resolves to a directory fails to open for writing on every Unix
        // platform, even as root
        let unrecoverable = temp_dir.path().join("bad.feather.partial");
        std::os::unix::fs::symlink(&run, &unrecoverable).unwrap();

        let warnings = recover_with_warnings(temp_dir.path());

        let sealed = partial.with_extension("");
        assert_eq!(read_quotes(&sealed), quotes_at(&[1]));
        assert_eq!(
            warnings,
            vec![
                format!(
                    "Failed to recover Feather file {}: Is a directory (os error 21)",
                    unrecoverable.display()
                ),
                format!(
                    "Recovered 1 record batch(es) into Feather file {}, dropped 8 trailing byte(s)",
                    sealed.display()
                ),
            ],
        );
    }

    #[rstest]
    fn recovery_leaves_open_and_empty_files() {
        let temp_dir = TempDir::new().unwrap();

        let mut writer = FeatherWriter::new(
            temp_dir.path().join("open"),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            Some(0),
        );
        writer.write(quote_at(1)).unwrap();
        writer.flush().unwrap();

        // Windows locks block reads through other handles, so the open file is compared by length
        let open = feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION).remove(0);
        let open_len = fs::metadata(&open).unwrap().len();
        let empty = temp_dir.path().join("empty.feather.partial");
        File::create(&empty).unwrap();

        recover_partial_feather_files(temp_dir.path());

        let partial = feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION);
        let open_len_after = fs::metadata(&open).unwrap().len();
        writer.close().unwrap();

        assert_eq!(partial, vec![empty, open.clone()]);
        assert_eq!(open_len_after, open_len);
        assert_eq!(read_quotes(&open.with_extension("")), quotes_at(&[1]));
    }

    #[rstest]
    fn size_rotation_in_mixed_identifier_batch_seals_the_shared_file_once() {
        let temp_dir = TempDir::new().unwrap();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::Size { max_size: 1 },
            None,
            None,
        );

        let quotes = ["AUD/USD.SIM", "EUR/USD.SIM"]
            .map(|instrument_id| {
                QuoteTick::new(
                    InstrumentId::from(instrument_id),
                    Price::from("1.0"),
                    Price::from("1.1"),
                    Quantity::from("1000"),
                    Quantity::from("1000"),
                    UnixNanos::from(1),
                    UnixNanos::from(1),
                )
            })
            .to_vec();

        // Both identifier groups fill the shared quotes file, so the batch queues its seal twice
        writer.write_batch(quotes).unwrap();

        let sealed = feather_files(temp_dir.path(), FEATHER_EXTENSION);
        let batches = read_feather_batches(&sealed[0]);

        assert!(writer.writers.is_empty());
        assert_eq!(sealed.len(), 1);
        assert!(feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION).is_empty());
        assert_eq!(
            identifiers_from_record_batches(&batches).unwrap(),
            vec!["AUD/USD.SIM".to_string(), "EUR/USD.SIM".to_string()],
        );
    }

    #[rstest]
    fn size_rotation_seals_each_full_file_and_opens_the_next_lazily() {
        let temp_dir = TempDir::new().unwrap();
        let shared_clock = Arc::new(AtomicU64::new(0));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::clone(&shared_clock)),
            RotationConfig::Size { max_size: 1 },
            None,
            Some(1),
        );

        for ts in [1_000, 2_000] {
            shared_clock.store(ts * 1_000_000, Ordering::Relaxed);
            writer
                .write(QuoteTick::new(
                    InstrumentId::from("AUD/USD.SIM"),
                    Price::from("1.0"),
                    Price::from("1.0"),
                    Quantity::from("1000"),
                    Quantity::from("1000"),
                    UnixNanos::from(ts),
                    UnixNanos::from(ts),
                ))
                .unwrap();
        }

        let rows = feather_files(temp_dir.path(), FEATHER_EXTENSION)
            .iter()
            .map(|path| {
                read_feather_batches(path)
                    .iter()
                    .map(RecordBatch::num_rows)
                    .sum::<usize>()
            })
            .collect::<Vec<_>>();

        assert!(writer.writers.is_empty());
        assert_eq!(rows, vec![1, 1]);
        assert!(feather_files(temp_dir.path(), FEATHER_PARTIAL_EXTENSION).is_empty());
    }

    #[rstest]
    #[case(
        "FeatherMissingTimestamp",
        Schema::empty(),
        "registered without an Arrow schema containing ts_init"
    )]
    #[case(
        "FeatherLegacyTimestamp",
        Schema::new(vec![Field::new("ts_init", DataType::UInt64, false)]),
        "registered with ts_init as UInt64",
    )]
    fn test_custom_writer_path_rejects_unqueryable_schema(
        #[case] type_name: &str,
        #[case] schema: Schema,
        #[case] expected_error: &str,
    ) {
        nautilus_model::data::registry::ensure_arrow_registered(
            type_name,
            Arc::new(schema),
            Box::new(|_| unreachable!("writer creation does not encode data")),
            Box::new(|_, _| unreachable!("writer creation does not decode data")),
        )
        .unwrap();

        let temp_dir = TempDir::new().unwrap();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            None,
        );

        for _ in 0..2 {
            let error = writer.get_writer_path_custom(type_name).unwrap_err();

            assert!(error.to_string().contains(expected_error));
            assert!(writer.writers.is_empty());
            assert!(writer.path_sequences.is_empty());
        }
    }

    #[rstest]
    #[cfg(feature = "python")]
    fn test_write_custom_data_round_trip() {
        use std::sync::Arc;

        use nautilus_model::{
            data::{CustomData, Data, DataType},
            identifiers::InstrumentId,
        };
        use nautilus_serialization::{
            arrow::custom::CustomDataDecoder, ensure_custom_data_registered,
        };

        use crate::test_data::RustTestCustomData;

        ensure_custom_data_registered::<RustTestCustomData>();

        let temp_dir = TempDir::new().unwrap();
        let clock = WriterClock::Test(Arc::new(AtomicU64::new(0)));

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            clock,
            RotationConfig::NoRotation,
            None,
            None,
        );

        let instrument_id = InstrumentId::from("RUST.TEST");
        let data_type = DataType::new("RustTestCustomData", None, Some(instrument_id.to_string()));

        let original = RustTestCustomData {
            instrument_id,
            value: 1.23,
            flag: true,
            ts_event: UnixNanos::from(1000),
            ts_init: UnixNanos::from(1000),
        };

        let custom = CustomData::new(Arc::new(original.clone()), data_type);

        writer
            .write_data(Data::Custom(custom))
            .expect("write_data CustomData");
        writer.close().expect("close");

        let prefix = temp_dir
            .path()
            .join("data")
            .join("custom")
            .join("RustTestCustomData");
        let files = feather_files(&prefix, FEATHER_EXTENSION);
        let batch = read_feather_batches(&files[0]).remove(0);
        let metadata = batch.schema().metadata().clone();
        let decoded =
            CustomDataDecoder::decode_data_batch(&metadata, batch).expect("decode_data_batch");
        assert_eq!(decoded.len(), 1);

        if let Data::Custom(decoded_custom) = &decoded[0] {
            assert_eq!(decoded_custom.data_type.type_name(), "RustTestCustomData");
            let rust: &RustTestCustomData = decoded_custom
                .data
                .as_any()
                .downcast_ref::<RustTestCustomData>()
                .expect("RustTestCustomData");
            assert_eq!(rust, &original);
        } else {
            panic!("Expected Data::Custom");
        }
    }

    #[rstest]
    #[cfg(feature = "python")]
    fn test_write_custom_data_reuses_writer_until_rotation() {
        use nautilus_model::data::{CustomData, DataType};
        use nautilus_serialization::ensure_custom_data_registered;

        use crate::test_data::RustTestCustomData;

        ensure_custom_data_registered::<RustTestCustomData>();
        let temp_dir = TempDir::new().unwrap();

        let mut writer = FeatherWriter::new(
            temp_dir.path().to_path_buf(),
            WriterClock::Test(Arc::new(AtomicU64::new(0))),
            RotationConfig::NoRotation,
            None,
            None,
        );
        let instrument_id = InstrumentId::from("RUST.TEST");
        let data_type = DataType::new("RustTestCustomData", None, Some(instrument_id.to_string()));

        for (ts, value) in [(1_000, 1_000.0), (2_000, 2_000.0)] {
            writer
                .write_data(Data::Custom(CustomData::new(
                    Arc::new(RustTestCustomData {
                        instrument_id,
                        value,
                        flag: true,
                        ts_event: UnixNanos::from(ts),
                        ts_init: UnixNanos::from(ts),
                    }),
                    data_type.clone(),
                )))
                .unwrap();
        }

        assert_eq!(writer.get_current_file_info().len(), 1);
        writer.close().unwrap();

        let prefix = temp_dir
            .path()
            .join("data")
            .join("custom")
            .join("RustTestCustomData");
        let files = feather_files(&prefix, FEATHER_EXTENSION);
        assert_eq!(files.len(), 1);
        let rows = read_feather_batches(&files[0])
            .iter()
            .map(RecordBatch::num_rows)
            .sum::<usize>();
        assert_eq!(rows, 2);
    }
}