perf(historian): group-commit batch append on the historization outbox
This commit is contained in:
@@ -19,6 +19,17 @@ public interface IHistorizationOutbox : IDisposable
|
||||
/// <returns>A task that represents the asynchronous operation.</returns>
|
||||
ValueTask AppendAsync(HistorizationOutboxEntry entry, CancellationToken ct);
|
||||
|
||||
/// <summary>
|
||||
/// Appends a burst of entries with ONE durability commit for the whole batch (group
|
||||
/// commit). Under PerEntry commit mode this is the per-batch fsync that replaces N
|
||||
/// per-entry fsyncs; under Periodic it is indistinguishable from N AppendAsync calls.
|
||||
/// An empty list is a no-op.
|
||||
/// </summary>
|
||||
/// <param name="entries">The value records to buffer, in append (FIFO) order.</param>
|
||||
/// <param name="ct">Cancellation token.</param>
|
||||
/// <returns>A task that represents the asynchronous operation.</returns>
|
||||
ValueTask AppendBatchAsync(IReadOnlyList<HistorizationOutboxEntry> entries, CancellationToken ct);
|
||||
|
||||
/// <summary>
|
||||
/// Returns up to <paramref name="max"/> oldest un-acked entries in FIFO order without removing
|
||||
/// them. Removal happens via <see cref="RemoveAsync"/> once each entry is durably written.
|
||||
|
||||
+37
-8
@@ -99,23 +99,52 @@ public sealed class FasterLogHistorizationOutbox : IHistorizationOutbox
|
||||
{
|
||||
ArgumentNullException.ThrowIfNull(entry);
|
||||
|
||||
byte[] payload = HistorizationOutboxEntrySerializer.Serialize(entry);
|
||||
long startAddress = await _log.EnqueueAsync(payload, ct).ConfigureAwait(false);
|
||||
// A one-entry batch is byte-for-byte the single-append path (one enqueue, one PerEntry
|
||||
// commit, one bookkeeping pass), so there is no second copy of the drop-oldest policy.
|
||||
await AppendBatchAsync([entry], ct).ConfigureAwait(false);
|
||||
}
|
||||
|
||||
/// <inheritdoc />
|
||||
public async ValueTask AppendBatchAsync(IReadOnlyList<HistorizationOutboxEntry> entries, CancellationToken ct)
|
||||
{
|
||||
ArgumentNullException.ThrowIfNull(entries);
|
||||
if (entries.Count == 0)
|
||||
return;
|
||||
|
||||
// Validate before ANY enqueue: a null halfway through would otherwise leave the batch's
|
||||
// earlier entries in the log with no in-memory index entry (invisible but not reclaimable).
|
||||
for (int i = 0; i < entries.Count; i++)
|
||||
ArgumentNullException.ThrowIfNull(entries[i]);
|
||||
|
||||
long[] startAddresses = new long[entries.Count];
|
||||
for (int i = 0; i < entries.Count; i++)
|
||||
{
|
||||
byte[] payload = HistorizationOutboxEntrySerializer.Serialize(entries[i]);
|
||||
startAddresses[i] = await _log.EnqueueAsync(payload, ct).ConfigureAwait(false);
|
||||
}
|
||||
|
||||
if (_commitMode == HistorizationCommitMode.PerEntry)
|
||||
{
|
||||
// PerEntry: durable before returning. Periodic skips this — the background timer (and
|
||||
// Dispose) commit on their cadence (accepted throughput/latency trade-off).
|
||||
// Group commit: ONE fsync makes the whole batch durable before returning, replacing the
|
||||
// N per-entry fsyncs a loop of AppendAsync would have cost. Periodic skips this — the
|
||||
// background timer (and Dispose) commit on their cadence (accepted trade-off).
|
||||
await _log.CommitAsync(ct).ConfigureAwait(false);
|
||||
}
|
||||
|
||||
long? truncateTo = null;
|
||||
lock (_state)
|
||||
{
|
||||
LinkedListNode<LiveEntry> node = _live.AddLast(new LiveEntry(entry.Id, startAddress));
|
||||
_index[entry.Id] = node;
|
||||
for (int i = 0; i < entries.Count; i++)
|
||||
{
|
||||
HistorizationOutboxEntry entry = entries[i];
|
||||
LinkedListNode<LiveEntry> node = _live.AddLast(new LiveEntry(entry.Id, startAddresses[i]));
|
||||
_index[entry.Id] = node;
|
||||
}
|
||||
|
||||
// Drop-oldest on overflow. The new head is the start address of whatever entry survives
|
||||
// at the front (or the tail if the log emptied); truncate to the furthest such address.
|
||||
// Drop-oldest on overflow, run ONCE for the whole batch (an over-capacity batch sheds
|
||||
// every surplus entry here). The new head is the start address of whatever entry
|
||||
// survives at the front (or the tail if the log emptied); truncate to the furthest such
|
||||
// address, so the batch needs at most one truncate + one extra commit.
|
||||
while (_capacity > 0 && _live.Count > _capacity)
|
||||
{
|
||||
LinkedListNode<LiveEntry> oldest = _live.First!;
|
||||
|
||||
+56
@@ -140,6 +140,62 @@ public sealed class FasterLogHistorizationOutboxTests : IDisposable
|
||||
Assert.Equal(new[] { a.Id, b.Id }, batch.Select(e => e.Id));
|
||||
}
|
||||
|
||||
[Fact]
|
||||
public async Task AppendBatch_then_peek_returns_fifo_in_batch_order()
|
||||
{
|
||||
// Group commit: one fsync for the whole batch. The reopen leg is the load-bearing half — it
|
||||
// proves the single per-batch commit actually made every entry durable, not just the last one.
|
||||
var dir = NewTempDir();
|
||||
var a = E("A", 1);
|
||||
var b = E("B", 2);
|
||||
var c = E("C", 3);
|
||||
{
|
||||
using var o = new FasterLogHistorizationOutbox(dir, HistorizationCommitMode.PerEntry);
|
||||
await o.AppendBatchAsync([a, b, c], TestContext.Current.CancellationToken);
|
||||
|
||||
Assert.Equal(3, await o.CountAsync(TestContext.Current.CancellationToken));
|
||||
var batch = await o.PeekBatchAsync(10, TestContext.Current.CancellationToken);
|
||||
Assert.Equal(new[] { a.Id, b.Id, c.Id }, batch.Select(e => e.Id));
|
||||
}
|
||||
|
||||
using var reopened = new FasterLogHistorizationOutbox(dir, HistorizationCommitMode.PerEntry);
|
||||
Assert.Equal(3, await reopened.CountAsync(TestContext.Current.CancellationToken));
|
||||
var survived = await reopened.PeekBatchAsync(10, TestContext.Current.CancellationToken);
|
||||
Assert.Equal(new[] { a.Id, b.Id, c.Id }, survived.Select(e => e.Id));
|
||||
}
|
||||
|
||||
[Fact]
|
||||
public async Task AppendBatch_at_capacity_drops_oldest_across_the_whole_batch()
|
||||
{
|
||||
// The drop-oldest loop runs ONCE for the batch, so an over-capacity batch must shed every
|
||||
// surplus entry (not just one) and count each drop.
|
||||
var dir = NewTempDir();
|
||||
using var o = new FasterLogHistorizationOutbox(dir, HistorizationCommitMode.PerEntry, capacity: 2);
|
||||
var a = E("A", 1);
|
||||
var b = E("B", 2);
|
||||
var c = E("C", 3);
|
||||
var d = E("D", 4);
|
||||
await o.AppendBatchAsync([a, b, c, d], TestContext.Current.CancellationToken);
|
||||
|
||||
Assert.Equal(2, await o.CountAsync(TestContext.Current.CancellationToken));
|
||||
Assert.Equal(2, o.DroppedCount);
|
||||
var tags = (await o.PeekBatchAsync(10, TestContext.Current.CancellationToken)).Select(e => e.Tag).ToArray();
|
||||
Assert.Equal(new[] { "C", "D" }, tags);
|
||||
}
|
||||
|
||||
[Fact]
|
||||
public async Task AppendBatch_empty_is_noop()
|
||||
{
|
||||
var dir = NewTempDir();
|
||||
using var o = new FasterLogHistorizationOutbox(dir, HistorizationCommitMode.PerEntry);
|
||||
await o.AppendAsync(E("A", 1), TestContext.Current.CancellationToken);
|
||||
|
||||
await o.AppendBatchAsync([], TestContext.Current.CancellationToken);
|
||||
|
||||
Assert.Equal(1, await o.CountAsync(TestContext.Current.CancellationToken));
|
||||
Assert.Equal(0, o.DroppedCount);
|
||||
}
|
||||
|
||||
[Fact]
|
||||
public async Task Remove_unknown_id_is_noop()
|
||||
{
|
||||
|
||||
+31
@@ -733,6 +733,37 @@ public sealed class ContinuousHistorizationRecorderTests : TestKit
|
||||
return ValueTask.CompletedTask;
|
||||
}
|
||||
|
||||
public ValueTask AppendBatchAsync(IReadOnlyList<HistorizationOutboxEntry> entries, CancellationToken ct)
|
||||
{
|
||||
if (entries.Count == 0)
|
||||
{
|
||||
return ValueTask.CompletedTask;
|
||||
}
|
||||
|
||||
if (_yieldOnAppend)
|
||||
{
|
||||
return AppendBatchYieldingAsync(entries);
|
||||
}
|
||||
|
||||
foreach (HistorizationOutboxEntry entry in entries)
|
||||
{
|
||||
AppendCore(entry);
|
||||
}
|
||||
|
||||
return ValueTask.CompletedTask;
|
||||
}
|
||||
|
||||
private async ValueTask AppendBatchYieldingAsync(IReadOnlyList<HistorizationOutboxEntry> entries)
|
||||
{
|
||||
// Same pool-thread-completion requirement as AppendYieldingAsync below — a group commit
|
||||
// fsyncs once for the batch, so the yield happens once, before the entries are recorded.
|
||||
await Task.Delay(1).ConfigureAwait(false);
|
||||
foreach (HistorizationOutboxEntry entry in entries)
|
||||
{
|
||||
AppendCore(entry);
|
||||
}
|
||||
}
|
||||
|
||||
private async ValueTask AppendYieldingAsync(HistorizationOutboxEntry entry)
|
||||
{
|
||||
// Must complete on a POOL thread, carrying no synchronization context — that is what a real
|
||||
|
||||
@@ -245,6 +245,9 @@ public sealed class ServiceCollectionExtensionsTests
|
||||
public long DroppedCount => 0;
|
||||
/// <summary>No-op append.</summary>
|
||||
public ValueTask AppendAsync(HistorizationOutboxEntry entry, CancellationToken ct) => ValueTask.CompletedTask;
|
||||
/// <summary>No-op batch append.</summary>
|
||||
public ValueTask AppendBatchAsync(IReadOnlyList<HistorizationOutboxEntry> entries, CancellationToken ct)
|
||||
=> ValueTask.CompletedTask;
|
||||
/// <summary>Always returns an empty batch.</summary>
|
||||
public ValueTask<IReadOnlyList<HistorizationOutboxEntry>> PeekBatchAsync(int max, CancellationToken ct)
|
||||
=> ValueTask.FromResult<IReadOnlyList<HistorizationOutboxEntry>>(Array.Empty<HistorizationOutboxEntry>());
|
||||
|
||||
Reference in New Issue
Block a user