s2_sdk/session/

append.rs

use std::{
    collections::VecDeque,
    future::Future,
    num::NonZeroU32,
    pin::Pin,
    sync::Arc,
    task::{Context, Poll},
    time::Duration,
};

use futures::StreamExt;
use tokio::{
    sync::{OwnedSemaphorePermit, Semaphore, mpsc, oneshot},
    time::Instant,
};
use tokio_muxt::{CoalesceMode, MuxTimer};
use tokio_stream::wrappers::ReceiverStream;
use tokio_util::task::AbortOnDropHandle;
use tracing::debug;

use crate::{
    api::{ApiError, BasinClient, Streaming, retry_builder},
    retry::RetryBackoffBuilder,
    types::{
        AppendAck, AppendInput, AppendRetryPolicy, MeteredBytes, ONE_MIB, S2Error, StreamName,
        StreamPosition, ValidationError,
    },
};

#[derive(Debug, thiserror::Error)]
pub enum AppendSessionError {
    #[error(transparent)]
    Api(#[from] ApiError),
    #[error("append acknowledgement timed out")]
    AckTimeout,
    #[error("server disconnected")]
    ServerDisconnected,
    #[error("response stream closed early while appends in flight")]
    StreamClosedEarly,
    #[error("session already closed")]
    SessionClosed,
    #[error("session is closing")]
    SessionClosing,
    #[error("session dropped without calling close")]
    SessionDropped,
}

impl AppendSessionError {
    pub fn is_retryable(&self) -> bool {
        match self {
            Self::Api(err) => err.is_retryable(),
            Self::AckTimeout => true,
            Self::ServerDisconnected => true,
            _ => false,
        }
    }
}

impl From<AppendSessionError> for S2Error {
    fn from(err: AppendSessionError) -> Self {
        match err {
            AppendSessionError::Api(api_err) => api_err.into(),
            other => S2Error::Client(other.to_string()),
        }
    }
}

/// A [`Future`] that resolves to an acknowledgement once the batch of records is appended.
pub struct BatchSubmitTicket {
    rx: oneshot::Receiver<Result<AppendAck, S2Error>>,
}

impl Future for BatchSubmitTicket {
    type Output = Result<AppendAck, S2Error>;

    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        match Pin::new(&mut self.rx).poll(cx) {
            Poll::Ready(Ok(res)) => Poll::Ready(res),
            Poll::Ready(Err(_)) => Poll::Ready(Err(AppendSessionError::SessionDropped.into())),
            Poll::Pending => Poll::Pending,
        }
    }
}

#[derive(Debug, Clone)]
/// Configuration for an [`AppendSession`].
pub struct AppendSessionConfig {
    max_inflight_bytes: u32,
    max_inflight_batches: Option<u32>,
}

impl Default for AppendSessionConfig {
    fn default() -> Self {
        Self {
            max_inflight_bytes: 10 * ONE_MIB,
            max_inflight_batches: None,
        }
    }
}

impl AppendSessionConfig {
    /// Create a new [`AppendSessionConfig`] with default settings.
    pub fn new() -> Self {
        Self::default()
    }

    /// Set the limit on total metered bytes of unacknowledged [`AppendInput`]s held in memory.
    ///
    /// **Note:** It must be at least `1MiB`.
    ///
    /// Defaults to `10MiB`.
    pub fn with_max_inflight_bytes(self, max_inflight_bytes: u32) -> Result<Self, ValidationError> {
        if max_inflight_bytes < ONE_MIB {
            return Err(format!("max_inflight_bytes must be at least {ONE_MIB}").into());
        }
        Ok(Self {
            max_inflight_bytes,
            ..self
        })
    }

    /// Set the limit on number of unacknowledged [`AppendInput`]s held in memory.
    ///
    /// Defaults to no limit.
    pub fn with_max_inflight_batches(
        self,
        max_inflight_batches: NonZeroU32,
    ) -> Result<Self, ValidationError> {
        Ok(Self {
            max_inflight_batches: Some(max_inflight_batches.get()),
            ..self
        })
    }
}

struct SessionState {
    cmd_rx: mpsc::Receiver<Command>,
    inflight_appends: VecDeque<InflightAppend>,
    inflight_bytes: usize,
    close_tx: Option<oneshot::Sender<Result<(), S2Error>>>,
    closing: bool,
    total_records: usize,
    total_acked_records: usize,
    prev_ack_end: Option<StreamPosition>,
}

/// A session for high-throughput appending with backpressure control. It can be created from
/// [`append_session`](crate::S2Stream::append_session).
///
/// Supports pipelining multiple [`AppendInput`]s while preserving submission order.
pub struct AppendSession {
    cmd_tx: mpsc::Sender<Command>,
    permits: AppendPermits,
    _handle: AbortOnDropHandle<()>,
}

impl AppendSession {
    pub(crate) fn new(
        client: BasinClient,
        stream: StreamName,
        config: AppendSessionConfig,
    ) -> Self {
        let buffer_size = config
            .max_inflight_batches
            .map(|mib| mib as usize)
            .unwrap_or(DEFAULT_CHANNEL_BUFFER_SIZE);
        let (cmd_tx, cmd_rx) = mpsc::channel(buffer_size);
        let permits = AppendPermits::new(config.max_inflight_batches, config.max_inflight_bytes);
        let retry_builder = retry_builder(&client.config.retry);
        let handle = AbortOnDropHandle::new(tokio::spawn(run_session_with_retry(
            client,
            stream,
            cmd_rx,
            retry_builder,
            buffer_size,
        )));
        Self {
            cmd_tx,
            permits,
            _handle: handle,
        }
    }

    /// Submit a batch of records for appending.
    ///
    /// Internally, it waits on [`reserve`](Self::reserve), then submits using the permit.
    /// This provides backpressure when inflight limits are reached.
    /// For explicit control, use [`reserve`](Self::reserve) followed by
    /// [`BatchSubmitPermit::submit`].
    ///
    /// **Note**: After all submits, you must call [`close`](Self::close) to ensure all batches are
    /// appended.
    pub async fn submit(&self, input: AppendInput) -> Result<BatchSubmitTicket, S2Error> {
        let permit = self.reserve(input.records.metered_bytes() as u32).await?;
        Ok(permit.submit(input))
    }

    /// Reserve capacity for a batch to be submitted. Useful in [`select!`](tokio::select) loops
    /// where you want to interleave submission with other async work. See [`submit`](Self::submit)
    /// for a simpler API.
    ///
    /// Waits when inflight limits are reached, providing explicit backpressure control.
    /// The returned permit must be used to submit the batch.
    ///
    /// **Note**: After all submits, you must call [`close`](Self::close) to ensure all batches are
    /// appended.
    ///
    /// # Cancel safety
    ///
    /// This method is cancel safe. Internally, it only awaits
    /// [`Semaphore::acquire_many_owned`](tokio::sync::Semaphore::acquire_many_owned) and
    /// [`Sender::reserve_owned`](tokio::sync::mpsc::Sender::reserve), both of which are cancel
    /// safe.
    pub async fn reserve(&self, bytes: u32) -> Result<BatchSubmitPermit, S2Error> {
        let append_permit = self.permits.acquire(bytes).await;
        let cmd_tx_permit = self
            .cmd_tx
            .clone()
            .reserve_owned()
            .await
            .map_err(|_| AppendSessionError::SessionClosed)?;
        Ok(BatchSubmitPermit {
            append_permit,
            cmd_tx_permit,
        })
    }

    /// Close the session and wait for all submitted batch of records to be appended.
    pub async fn close(self) -> Result<(), S2Error> {
        let (done_tx, done_rx) = oneshot::channel();
        self.cmd_tx
            .send(Command::Close { done_tx })
            .await
            .map_err(|_| AppendSessionError::SessionClosed)?;
        done_rx
            .await
            .map_err(|_| AppendSessionError::SessionClosed)??;
        Ok(())
    }
}

/// A permit to submit a batch after reserving capacity.
pub struct BatchSubmitPermit {
    append_permit: AppendPermit,
    cmd_tx_permit: mpsc::OwnedPermit<Command>,
}

impl BatchSubmitPermit {
    /// Submit the batch using this permit.
    pub fn submit(self, input: AppendInput) -> BatchSubmitTicket {
        let (ack_tx, ack_rx) = oneshot::channel();
        self.cmd_tx_permit.send(Command::Submit {
            input,
            ack_tx,
            permit: Some(self.append_permit),
        });
        BatchSubmitTicket { rx: ack_rx }
    }
}

pub(crate) struct AppendSessionInternal {
    cmd_tx: mpsc::Sender<Command>,
    _handle: AbortOnDropHandle<()>,
}

impl AppendSessionInternal {
    pub(crate) fn new(client: BasinClient, stream: StreamName) -> Self {
        let buffer_size = DEFAULT_CHANNEL_BUFFER_SIZE;
        let (cmd_tx, cmd_rx) = mpsc::channel(buffer_size);
        let retry_builder = retry_builder(&client.config.retry);
        let handle = AbortOnDropHandle::new(tokio::spawn(run_session_with_retry(
            client,
            stream,
            cmd_rx,
            retry_builder,
            buffer_size,
        )));
        Self {
            cmd_tx,
            _handle: handle,
        }
    }

    pub(crate) fn submit(
        &self,
        input: AppendInput,
    ) -> impl Future<Output = Result<BatchSubmitTicket, S2Error>> + Send + 'static {
        let cmd_tx = self.cmd_tx.clone();
        async move {
            let (ack_tx, ack_rx) = oneshot::channel();
            cmd_tx
                .send(Command::Submit {
                    input,
                    ack_tx,
                    permit: None,
                })
                .await
                .map_err(|_| AppendSessionError::SessionClosed)?;
            Ok(BatchSubmitTicket { rx: ack_rx })
        }
    }

    pub(crate) async fn close(self) -> Result<(), S2Error> {
        let (done_tx, done_rx) = oneshot::channel();
        self.cmd_tx
            .send(Command::Close { done_tx })
            .await
            .map_err(|_| AppendSessionError::SessionClosed)?;
        done_rx
            .await
            .map_err(|_| AppendSessionError::SessionClosed)??;
        Ok(())
    }
}

#[derive(Debug)]
pub(crate) struct AppendPermit {
    _count: Option<OwnedSemaphorePermit>,
    _bytes: OwnedSemaphorePermit,
}

#[derive(Clone)]
pub(crate) struct AppendPermits {
    count: Option<Arc<Semaphore>>,
    bytes: Arc<Semaphore>,
}

impl AppendPermits {
    pub(crate) fn new(count_permits: Option<u32>, bytes_permits: u32) -> Self {
        Self {
            count: count_permits.map(|permits| Arc::new(Semaphore::new(permits as usize))),
            bytes: Arc::new(Semaphore::new(bytes_permits as usize)),
        }
    }

    pub(crate) async fn acquire(&self, bytes: u32) -> AppendPermit {
        AppendPermit {
            _count: if let Some(count) = self.count.as_ref() {
                Some(
                    count
                        .clone()
                        .acquire_many_owned(1)
                        .await
                        .expect("semaphore should not be closed"),
                )
            } else {
                None
            },
            _bytes: self
                .bytes
                .clone()
                .acquire_many_owned(bytes)
                .await
                .expect("semaphore should not be closed"),
        }
    }
}

async fn run_session_with_retry(
    client: BasinClient,
    stream: StreamName,
    cmd_rx: mpsc::Receiver<Command>,
    retry_builder: RetryBackoffBuilder,
    buffer_size: usize,
) {
    let mut state = SessionState {
        cmd_rx,
        inflight_appends: VecDeque::new(),
        inflight_bytes: 0,
        close_tx: None,
        closing: false,
        total_records: 0,
        total_acked_records: 0,
        prev_ack_end: None,
    };
    let mut prev_total_acked_records = 0;
    let mut retry_backoffs = retry_builder.build();

    loop {
        let result = run_session(&client, &stream, &mut state, buffer_size).await;

        match result {
            Ok(()) => {
                break;
            }
            Err(err) => {
                if prev_total_acked_records < state.total_acked_records {
                    prev_total_acked_records = state.total_acked_records;
                    retry_backoffs.reset();
                }

                let retry_policy_compliant = retry_policy_compliant(
                    client.config.retry.append_retry_policy,
                    &state.inflight_appends,
                );

                if retry_policy_compliant
                    && err.is_retryable()
                    && let Some(backoff) = retry_backoffs.next()
                {
                    debug!(
                        %err,
                        ?backoff,
                        num_retries_remaining = retry_backoffs.remaining(),
                        "retrying append session"
                    );
                    tokio::time::sleep(backoff).await;
                } else {
                    debug!(
                        %err,
                        retry_policy_compliant,
                        retries_exhausted = retry_backoffs.is_exhausted(),
                        "not retrying append session"
                    );

                    let err: S2Error = err.into();
                    for inflight_append in state.inflight_appends.drain(..) {
                        let _ = inflight_append.ack_tx.send(Err(err.clone()));
                    }

                    if let Some(done_tx) = state.close_tx.take() {
                        let _ = done_tx.send(Err(err));
                    }
                    break;
                }
            }
        }
    }

    if let Some(done_tx) = state.close_tx.take() {
        let _ = done_tx.send(Ok(()));
    }
}

async fn run_session(
    client: &BasinClient,
    stream: &StreamName,
    state: &mut SessionState,
    buffer_size: usize,
) -> Result<(), AppendSessionError> {
    let (input_tx, mut acks) = connect(client, stream, buffer_size).await?;
    let ack_timeout = client.config.request_timeout;

    if !state.inflight_appends.is_empty() {
        resend(state, &input_tx, &mut acks, ack_timeout).await?;

        assert!(state.inflight_appends.is_empty());
        assert_eq!(state.inflight_bytes, 0);
    }

    let timer = MuxTimer::<N_TIMER_VARIANTS>::default();
    tokio::pin!(timer);

    let mut stashed_submission: Option<StashedSubmission> = None;

    loop {
        tokio::select! {
            (event_ord, _deadline) = &mut timer, if timer.is_armed() => {
                match TimerEvent::from(event_ord) {
                    TimerEvent::AckDeadline => {
                        return Err(AppendSessionError::AckTimeout);
                    }
                }
            }

            input_tx_permit = input_tx.reserve(), if stashed_submission.is_some() => {
                let input_tx_permit = input_tx_permit
                    .map_err(|_| AppendSessionError::ServerDisconnected)?;
                let submission = stashed_submission
                    .take()
                    .expect("stashed_submission should not be None");

                input_tx_permit.send(submission.input.clone());

                state.total_records += submission.input.records.len();
                state.inflight_bytes += submission.input_metered_bytes;

                timer.as_mut().fire_at(
                    TimerEvent::AckDeadline,
                    submission.since + ack_timeout,
                    CoalesceMode::Earliest,
                );
                state.inflight_appends.push_back(submission.into());
            }

            cmd = state.cmd_rx.recv(), if stashed_submission.is_none() => {
                match cmd {
                    Some(Command::Submit { input, ack_tx, permit }) => {
                        if state.closing {
                            let _ = ack_tx.send(
                                Err(AppendSessionError::SessionClosing.into())
                            );
                        } else {
                            let input_metered_bytes = input.records.metered_bytes();
                            stashed_submission = Some(StashedSubmission {
                                input,
                                input_metered_bytes,
                                ack_tx,
                                permit,
                                since: Instant::now(),
                            });
                        }
                    }
                    Some(Command::Close { done_tx }) => {
                        state.closing = true;
                        state.close_tx = Some(done_tx);
                    }
                    None => {
                        return Err(AppendSessionError::SessionDropped);
                    }
                }
            }

            ack = acks.next() => {
                match ack {
                    Some(Ok(ack)) => {
                        process_ack(
                            ack,
                            state,
                            timer.as_mut(),
                            ack_timeout,
                        );
                    }
                    Some(Err(err)) => {
                        return Err(err.into());
                    }
                    None => {
                        if !state.inflight_appends.is_empty() || stashed_submission.is_some() {
                            return Err(AppendSessionError::StreamClosedEarly);
                        }
                        break;
                    }
                }
            }
        }

        if state.closing && state.inflight_appends.is_empty() && stashed_submission.is_none() {
            break;
        }
    }

    assert!(state.inflight_appends.is_empty());
    assert_eq!(state.inflight_bytes, 0);
    assert!(stashed_submission.is_none());

    Ok(())
}

async fn resend(
    state: &mut SessionState,
    input_tx: &mpsc::Sender<AppendInput>,
    acks: &mut Streaming<AppendAck>,
    ack_timeout: Duration,
) -> Result<(), AppendSessionError> {
    debug!(
        inflight_appends_len = state.inflight_appends.len(),
        inflight_bytes = state.inflight_bytes,
        "resending inflight appends"
    );

    let mut resend_index = 0;
    let mut resend_finished = false;

    let timer = MuxTimer::<N_TIMER_VARIANTS>::default();
    tokio::pin!(timer);

    while !state.inflight_appends.is_empty() {
        tokio::select! {
            (event_ord, _deadline) = &mut timer, if timer.is_armed() => {
                match TimerEvent::from(event_ord) {
                    TimerEvent::AckDeadline => {
                        return Err(AppendSessionError::AckTimeout);
                    }
                }
            }

            input_tx_permit = input_tx.reserve(), if !resend_finished => {
                let input_tx_permit = input_tx_permit
                    .map_err(|_| AppendSessionError::ServerDisconnected)?;

                if let Some(inflight_append) = state.inflight_appends.get_mut(resend_index) {
                    inflight_append.since = Instant::now();
                    timer.as_mut().fire_at(
                        TimerEvent::AckDeadline,
                        inflight_append.since + ack_timeout,
                        CoalesceMode::Earliest,
                    );
                    input_tx_permit.send(inflight_append.input.clone());
                    resend_index += 1;
                } else {
                    resend_finished = true;
                }
            }

            ack = acks.next() => {
                match ack {
                    Some(Ok(ack)) => {
                        process_ack(
                            ack,
                            state,
                            timer.as_mut(),
                            ack_timeout,
                        );
                        resend_index -= 1;
                    }
                    Some(Err(err)) => {
                        return Err(err.into());
                    }
                    None => {
                        return Err(AppendSessionError::StreamClosedEarly);
                    }
                }
            }
        }
    }

    assert_eq!(
        resend_index, 0,
        "resend_index should be 0 after resend completes"
    );
    debug!("finished resending inflight appends");
    Ok(())
}

async fn connect(
    client: &BasinClient,
    stream: &StreamName,
    buffer_size: usize,
) -> Result<(mpsc::Sender<AppendInput>, Streaming<AppendAck>), AppendSessionError> {
    let (input_tx, input_rx) = mpsc::channel::<AppendInput>(buffer_size);
    let ack_stream = Box::pin(
        client
            .append_session(stream, ReceiverStream::new(input_rx).map(|i| i.into()))
            .await?
            .map(|ack| match ack {
                Ok(ack) => Ok(ack.into()),
                Err(err) => Err(err),
            }),
    );
    Ok((input_tx, ack_stream))
}

fn process_ack(
    ack: AppendAck,
    state: &mut SessionState,
    timer: Pin<&mut MuxTimer<N_TIMER_VARIANTS>>,
    ack_timeout: Duration,
) {
    let corresponding_append = state
        .inflight_appends
        .pop_front()
        .expect("corresponding append should be present for an ack");

    assert!(
        ack.end.seq_num >= ack.start.seq_num,
        "ack end seq_num should be greater than or equal to start seq_num"
    );

    if let Some(end) = state.prev_ack_end {
        assert!(
            ack.end.seq_num > end.seq_num,
            "ack end seq_num should be greater than previous ack end"
        );
    }

    let num_acked_records = (ack.end.seq_num - ack.start.seq_num) as usize;
    assert_eq!(
        num_acked_records,
        corresponding_append.input.records.len(),
        "ack record count should match submitted batch size"
    );

    state.total_acked_records += num_acked_records;
    state.inflight_bytes -= corresponding_append.input_metered_bytes;
    state.prev_ack_end = Some(ack.end);

    let _ = corresponding_append.ack_tx.send(Ok(ack));

    if let Some(oldest_append) = state.inflight_appends.front() {
        timer.fire_at(
            TimerEvent::AckDeadline,
            oldest_append.since + ack_timeout,
            CoalesceMode::Latest,
        );
    } else {
        timer.cancel(TimerEvent::AckDeadline);
        assert_eq!(
            state.total_records, state.total_acked_records,
            "all records should be acked when inflight is empty"
        );
    }
}

struct StashedSubmission {
    input: AppendInput,
    input_metered_bytes: usize,
    ack_tx: oneshot::Sender<Result<AppendAck, S2Error>>,
    permit: Option<AppendPermit>,
    since: Instant,
}

struct InflightAppend {
    input: AppendInput,
    input_metered_bytes: usize,
    ack_tx: oneshot::Sender<Result<AppendAck, S2Error>>,
    since: Instant,
    _permit: Option<AppendPermit>,
}

impl From<StashedSubmission> for InflightAppend {
    fn from(value: StashedSubmission) -> Self {
        Self {
            input: value.input,
            input_metered_bytes: value.input_metered_bytes,
            ack_tx: value.ack_tx,
            since: value.since,
            _permit: value.permit,
        }
    }
}

fn retry_policy_compliant(
    policy: AppendRetryPolicy,
    inflight_appends: &VecDeque<InflightAppend>,
) -> bool {
    if policy == AppendRetryPolicy::All {
        return true;
    }
    inflight_appends
        .iter()
        .all(|ia| policy.is_compliant(&ia.input))
}

enum Command {
    Submit {
        input: AppendInput,
        ack_tx: oneshot::Sender<Result<AppendAck, S2Error>>,
        permit: Option<AppendPermit>,
    },
    Close {
        done_tx: oneshot::Sender<Result<(), S2Error>>,
    },
}

const DEFAULT_CHANNEL_BUFFER_SIZE: usize = 100;

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum TimerEvent {
    AckDeadline,
}

const N_TIMER_VARIANTS: usize = 1;

impl From<TimerEvent> for usize {
    fn from(event: TimerEvent) -> Self {
        match event {
            TimerEvent::AckDeadline => 0,
        }
    }
}

impl From<usize> for TimerEvent {
    fn from(value: usize) -> Self {
        match value {
            0 => TimerEvent::AckDeadline,
            _ => panic!("invalid ordinal"),
        }
    }
}