lora_wal/recorder/

recorder.rs

//! [`WalRecorder`] — adapter from `MutationRecorder` to the durable
//! [`Wal`].
//!
//! Lifecycle, viewed from `lora-database::Database::execute_with_params`:
//!
//! 1. Acquire the store write lock.
//! 2. `recorder.arm()` — marks the recorder as inside-a-query but
//!    appends nothing to the WAL yet. A pure read query that fires
//!    no `MutationEvent` therefore touches the WAL zero times.
//! 3. Run analyze + compile + execute. The executor mutates the
//!    in-memory store, which fires `MutationRecorder::record` for each
//!    primitive mutation. The adapter buffers those events in memory.
//! 4. On Ok: `recorder.commit()` drains the buffered events and hands
//!    them to [`Wal::commit_tx`], which writes `TxBegin` +
//!    `MutationBatch` + `TxCommit` in one critical section and applies
//!    the configured single-thread flush policy. A read-only query returns
//!    `WroteCommit::No` and the WAL never wakes up.
//! 5. On Err / panic: `recorder.abort()` discards the buffered events.
//!    Because `commit_tx` writes the begin/batch/commit triple
//!    atomically, an aborted query has *no* on-disk presence — there
//!    is no `TxBegin` to pair with a later `TxAbort`, so the WAL stays
//!    consistent without an explicit abort marker.
//! 6. Before returning, the host inspects `recorder.poisoned()` once.
//!    If `Some`, the query fails loudly with a durability error so
//!    the caller can act on it; the WAL is now refusing further
//!    appends until the operator restarts the database, which
//!    recovers from the last consistent snapshot + WAL.
//!
//! ### Hot-path cost
//!
//! `record` is called once per primitive mutation. It takes only the
//! recorder mutex and pushes the event into a query-local buffer; the
//! WAL mutex, framing, checksum, and segment append happen once at
//! commit time inside `Wal::commit_tx`.
//!
//! ### When `record` fires after a failed in-memory mutation
//!
//! `InMemoryGraph::emit` only calls the recorder *after* the mutation
//! has been committed to the in-memory state. If the subsequent WAL
//! append fails, the live in-memory store is briefly ahead of disk:
//! the next query sees the partial state, but the next query also
//! observes `poisoned() = Some(_)` and is rejected. Recovery from a
//! snapshot + WAL after operator restart will not include the failed
//! mutation, so durable state stays consistent. The cost is "the live
//! process is wrong until the next restart"; the gain is that the
//! storage trait does not need to learn about durability.

use std::sync::{Arc, Mutex, MutexGuard};

use lora_store::{MutationEvent, MutationRecorder};

use super::errors::{WalBufferedCommitError, WalPoisonError, WroteCommit};
use super::mirror::WalMirror;
use crate::errors::WalError;
use crate::lsn::Lsn;
use crate::wal::Wal;

#[derive(Default)]
struct RecorderState {
    /// True between `arm()` and the matching `commit()` / `abort()`.
    /// Marks the host's critical section so [`MutationRecorder::record`]
    /// knows whether to buffer fresh events or poison itself for an
    /// out-of-scope call.
    armed: bool,
    /// Query-local mutation buffer. Drained by `commit()` and passed
    /// to [`Wal::commit_tx`] as one batched WAL transaction; cleared
    /// by `abort()` without ever touching the durable log.
    buffer: Vec<MutationEvent>,
    /// Sticky failure flag. Once set, [`MutationRecorder::record`]
    /// becomes a no-op (we cannot append safely) and `poisoned`
    /// surfaces the message.
    poisoned: Option<String>,
}

/// Adapter that lets a [`Wal`] act as a [`MutationRecorder`] on
/// [`lora_store::InMemoryGraph::set_mutation_recorder`].
pub struct WalRecorder {
    wal: Arc<Wal>,
    mirror: Option<Arc<dyn WalMirror>>,
    state: Mutex<RecorderState>,
}

impl WalRecorder {
    pub fn new(wal: Arc<Wal>) -> Self {
        Self::new_with_mirror(wal, None)
    }

    pub fn new_with_mirror(wal: Arc<Wal>, mirror: Option<Arc<dyn WalMirror>>) -> Self {
        Self {
            wal,
            mirror,
            state: Mutex::new(RecorderState::default()),
        }
    }

    /// Underlying log handle. Exposed so admin paths
    /// (`Database::checkpoint_to`, `truncate_up_to`) can hit the WAL
    /// directly without going through the recorder's transaction
    /// state machine.
    pub fn wal(&self) -> &Arc<Wal> {
        &self.wal
    }

    fn state_lock(&self) -> MutexGuard<'_, RecorderState> {
        self.state
            .lock()
            .unwrap_or_else(|poisoned| poisoned.into_inner())
    }

    /// Mark the recorder as inside a query critical section. No WAL
    /// I/O happens here — `Wal::begin` is deferred until the first
    /// mutation event fires. A pure read query that never produces a
    /// `MutationEvent` therefore costs the WAL nothing: no record
    /// allocation, no buffer drain, no `fsync`.
    ///
    /// Errors with [`WalError::Poisoned`] if a prior failure has
    /// poisoned the recorder, or if the host is double-arming
    /// (`arm` already in effect).
    pub fn arm(&self) -> Result<(), WalError> {
        let mut state = self.state_lock();
        if state.poisoned.is_some() {
            return Err(WalError::Poisoned);
        }
        if state.armed {
            state.poisoned = Some("WalRecorder::arm called while already armed".into());
            return Err(WalError::Poisoned);
        }
        state.armed = true;
        state.buffer.clear();
        Ok(())
    }

    /// Drain the buffered events and commit them as one durable WAL
    /// transaction.
    ///
    /// Routes through [`Wal::commit_tx`], which encodes
    /// `TxBegin` + `MutationBatch` + `TxCommit` in a single critical
    /// section and applies the configured flush policy. Under
    /// `SyncMode::PerCommit`, durability is complete when this method returns
    /// — the legacy "commit then flush" split is gone.
    ///
    /// Returns:
    /// - [`WroteCommit::Yes`] when mutation events fired and the WAL
    ///   wrote the begin/batch/commit triple.
    /// - [`WroteCommit::No`] when no mutations fired during the query
    ///   and no records were written.
    pub fn commit(&self) -> Result<WroteCommit, WalError> {
        let events = {
            let mut state = self.state_lock();
            if state.poisoned.is_some() {
                return Err(WalError::Poisoned);
            }
            if !state.armed {
                state.poisoned = Some("WalRecorder::commit called without an armed query".into());
                return Err(WalError::Poisoned);
            }
            state.armed = false;
            std::mem::take(&mut state.buffer)
        };

        if events.is_empty() {
            return Ok(WroteCommit::No);
        }

        self.wal.commit_tx(events).inspect_err(|e| {
            self.state_lock()
                .poisoned
                .get_or_insert_with(|| e.to_string());
        })
    }

    /// Commit an explicit transaction's externally-buffered mutation
    /// events as one durable WAL transaction.
    ///
    /// Used by `lora-database`'s [`Transaction`] flow, which keeps its
    /// own `Vec<MutationEvent>` per transaction (statements may
    /// rollback to a savepoint, which the recorder's flat buffer
    /// cannot model). At commit time the host hands the accumulated
    /// events here and we route them through [`Wal::commit_tx`] in one
    /// call.
    ///
    /// [`Transaction`]: lora_database::Transaction
    pub fn commit_events(
        &self,
        events: impl IntoIterator<Item = MutationEvent>,
    ) -> Result<WroteCommit, WalBufferedCommitError> {
        self.ensure_not_poisoned()
            .map_err(|e| WalBufferedCommitError::Poisoned(e.reason().to_string()))?;

        let events: Vec<MutationEvent> = events.into_iter().collect();
        if events.is_empty() {
            return Ok(WroteCommit::No);
        }

        self.wal.commit_tx(events).map_err(|e| {
            self.state_lock()
                .poisoned
                .get_or_insert_with(|| e.to_string());
            WalBufferedCommitError::Commit(super::errors::WalCommitError::Commit(e))
        })
    }

    /// Discard any buffered events and disarm the recorder.
    ///
    /// Because [`Wal::commit_tx`] writes the entire begin/batch/commit
    /// triple atomically, an aborted query never has any on-disk
    /// presence — there is no half-written transaction to follow up
    /// with a `TxAbort` marker. The returned bool indicates whether
    /// the query observed any mutations (so the host can decide
    /// whether to quarantine the live in-memory graph).
    pub fn abort(&self) -> Result<bool, WalError> {
        let mut state = self.state_lock();
        if state.poisoned.is_some() {
            return Err(WalError::Poisoned);
        }
        // Tolerate "abort without arm" — the host calls abort in
        // unwind paths and we'd rather no-op than poison.
        state.armed = false;
        let had_buffered_events = !state.buffer.is_empty();
        state.buffer.clear();
        Ok(had_buffered_events)
    }

    /// Flush the WAL — write the pending buffer to the OS and
    /// (under `SyncMode::PerCommit`) `fsync`.
    pub fn flush(&self) -> Result<(), WalError> {
        let mut state = self.state_lock();
        if state.poisoned.is_some() {
            return Err(WalError::Poisoned);
        }
        self.wal.flush().inspect_err(|e| {
            state.poisoned = Some(e.to_string());
        })?;
        if let Some(mirror) = &self.mirror {
            mirror.persist(self.wal.dir()).inspect_err(|e| {
                state.poisoned = Some(e.to_string());
            })?;
        }
        Ok(())
    }

    /// Force the underlying WAL to write, `fsync`, and advance its
    /// durable fence regardless of the configured sync mode. Admin
    /// paths use this when they need a durability point immediately.
    pub fn force_fsync(&self) -> Result<(), WalError> {
        let mut state = self.state_lock();
        if state.poisoned.is_some() {
            return Err(WalError::Poisoned);
        }
        self.wal.force_fsync().inspect_err(|e| {
            state.poisoned = Some(e.to_string());
        })?;
        if let Some(mirror) = &self.mirror {
            mirror.persist_force(self.wal.dir()).inspect_err(|e| {
                state.poisoned = Some(e.to_string());
            })?;
        }
        Ok(())
    }

    /// Append a `Checkpoint` marker. Used by the checkpoint admin
    /// path after a successful snapshot rename — the marker doubles
    /// as the log-side fence the next replay will trust.
    pub fn checkpoint_marker(&self, snapshot_lsn: Lsn) -> Result<Lsn, WalError> {
        let mut state = self.state_lock();
        if state.poisoned.is_some() {
            return Err(WalError::Poisoned);
        }
        self.wal.checkpoint_marker(snapshot_lsn).inspect_err(|e| {
            state.poisoned = Some(e.to_string());
        })
    }

    /// Drop sealed segments at or below `fence_lsn`. Forwards to
    /// [`Wal::truncate_up_to`].
    pub fn truncate_up_to(&self, fence_lsn: Lsn) -> Result<(), WalError> {
        // Archive-backed databases must stay self-contained. Until snapshot
        // checkpoint payloads are stored inside the archive too, preserving the
        // full WAL history is the only safe way to let the archive recover by
        // itself after a checkpoint marker.
        if let Some(mirror) = &self.mirror {
            mirror.persist_force(self.wal.dir())?;
            return Ok(());
        }
        self.wal.truncate_up_to(fence_lsn)?;
        Ok(())
    }

    /// True iff the recorder has already failed an append, **or** the WAL has
    /// latched a durability failure. Cheap to poll under the store lock.
    pub fn is_poisoned(&self) -> bool {
        self.poisoned_reason().is_some()
    }

    pub fn poisoned_reason(&self) -> Option<String> {
        let state = self.state_lock();
        if let Some(msg) = state.poisoned.clone() {
            return Some(msg);
        }
        self.wal.bg_failure()
    }

    pub fn ensure_not_poisoned(&self) -> Result<(), WalPoisonError> {
        if let Some(reason) = self.poisoned_reason() {
            return Err(WalPoisonError { reason });
        }
        Ok(())
    }

    /// Quarantine the recorder after the host detects that the live
    /// in-memory graph may no longer match durable state. Once poisoned,
    /// future query arms fail until the database is restarted from a
    /// snapshot + WAL.
    pub fn poison(&self, reason: impl Into<String>) {
        let mut state = self.state_lock();
        state.poisoned.get_or_insert_with(|| reason.into());
        state.armed = false;
        state.buffer.clear();
    }

    /// Test helper: clear the poisoned flag and disarm. Production
    /// code should not call this — once the WAL is poisoned the right
    /// move is to fail loudly and let the operator restart from the
    /// last snapshot + WAL.
    #[doc(hidden)]
    pub fn clear_poisoned_for_tests(&self) {
        let mut state = self.state_lock();
        state.poisoned = None;
        state.armed = false;
        state.buffer.clear();
    }
}

impl MutationRecorder for WalRecorder {
    fn record(&self, event: MutationEvent) {
        let mut state = self.state_lock();
        if state.poisoned.is_some() {
            return;
        }
        if !state.armed {
            state.poisoned.get_or_insert_with(|| {
                "MutationRecorder::record fired outside an armed query".into()
            });
            return;
        }
        state.buffer.push(event);
    }

    fn poisoned(&self) -> Option<String> {
        // Surface a latched WAL failure too — the recorder is the host's
        // single point of contact for "is the WAL still safe to commit
        // through?".
        self.poisoned_reason()
    }
}