Files
mxaccessgw/src/ZB.MOM.WW.MxGateway.Server/Grpc/EventStreamService.cs
T
Joseph Doherty 5e2e40a927 perf(gateway): trim event/command hot-path allocations (GWC-06/07/15, IPC-05)
Behavior-preserving allocation cuts on the per-event/per-command path:
- GWC-06: StreamEvents send-timing uses Stopwatch.GetTimestamp() +
  GetElapsedTime() instead of a per-event Stopwatch allocation (same measured
  span).
- GWC-07/IPC-05 (event): MapEvent transfers ownership of the inner MxEvent
  instead of .Clone()-ing it. Safe: WorkerEvent is parsed fresh per pipe frame
  with the distributor pump as its single consumer (GWC-01), MapEvent runs once
  before fan-out, and every downstream consumer (subscribers, replay ring)
  only READS the event (WorkerSequence is stamped upstream; verified no
  post-mapping mutation). Comment documents the invariant + restore-clone
  caveat if a second consumer is added.
- IPC-05 (command): CreateCommandEnvelope no longer re-clones; MapCommand
  already isolated the graph from the caller-owned gRPC message.
- GWC-15: grpc_stream_queue.depth converts from a per-event push counter to an
  ObservableGauge summing registered channel sources at scrape time only
  (name/semantics unchanged); removes all per-event .Count/lock work.

Kept every load-bearing isolation clone (MapCommand, Invoke, bulk filters,
MapCommandReply). Server build clean (0 warnings); EventStream/Metrics/
Distributor/Mapper tests 62/62 (incl. formerly-flaky queue-depth tests, now
green under the lazy gauge, + 2 new MapEvent ownership tests). Docs: Metrics.md,
Grpc.md.

Claude-Session: https://claude.ai/code/session_01DMXXvNuPekkkrTEyPNxEkW
2026-07-09 15:12:43 -04:00

234 lines
12 KiB
C#

using System.Runtime.CompilerServices;
using Microsoft.Extensions.Options;
using ZB.MOM.WW.MxGateway.Contracts.Proto;
using ZB.MOM.WW.MxGateway.Server.Configuration;
using ZB.MOM.WW.MxGateway.Server.Metrics;
using ZB.MOM.WW.MxGateway.Server.Sessions;
using ZB.MOM.WW.MxGateway.Server.Workers;
namespace ZB.MOM.WW.MxGateway.Server.Grpc;
public sealed class EventStreamService(
ISessionManager sessionManager,
IOptions<GatewayOptions> options,
GatewayMetrics metrics) : IEventStreamService
{
/// <inheritdoc />
/// <remarks>
/// <para>
/// This reads the subscriber's lease channel fed by the session's single
/// <see cref="SessionEventDistributor"/> pump. The pump owns the single drain of
/// the worker event stream and the worker→public mapping (mirroring the former
/// <c>ProduceEventsAsync</c>); this loop is the per-subscriber boundary that
/// applies the per-RPC filter (<c>AfterWorkerSequence</c>), queue-depth metrics,
/// and the backpressure/overflow policy.
/// </para>
/// <para>
/// The dashboard mirror runs OFF this per-RPC loop. The dashboard is a
/// first-class internal subscriber on the session's
/// <see cref="SessionEventDistributor"/> (see <c>GatewaySession.StartDashboardMirror</c>),
/// so it receives session events even when no gRPC client is streaming. This loop
/// does not mirror to the dashboard. One deliberate consequence: the dashboard sees
/// RAW session events, not the per-gRPC-subscriber <c>AfterWorkerSequence</c>-filtered
/// view this loop applies — the dashboard is a separate LDAP-authenticated monitoring
/// view that should see the session's full event activity.
/// </para>
/// <para>
/// Overflow handling: the distributor's per-subscriber channel is bounded
/// and the pump writes non-blocking. When this subscriber's channel is full the pump
/// applies the per-subscriber backpressure policy and completes this subscriber's
/// channel with a <see cref="SessionManagerException"/>
/// (<see cref="SessionManagerErrorCode.EventQueueOverflow"/>). That terminal fault
/// surfaces here when the reader's <c>MoveNextAsync</c> throws, and it propagates to
/// the gRPC client unchanged. The overflow metric, and (in the legacy
/// single-subscriber FailFast case) the session fault + fault metric, are recorded by
/// the distributor's overflow handler so the session, the pump, and other subscribers
/// are isolated from this subscriber's slowness.
/// </para>
/// </remarks>
public async IAsyncEnumerable<MxEvent> StreamEventsAsync(
StreamEventsRequest request,
string? callerKeyId,
[EnumeratorCancellation] CancellationToken cancellationToken)
{
if (!sessionManager.TryGetSession(request.SessionId, out GatewaySession? session) || session is null)
{
throw new SessionManagerException(
SessionManagerErrorCode.SessionNotFound,
$"Session {request.SessionId} was not found.");
}
// Owner-scoped attach (TST-02, security control): a session's event stream may be
// attached or reattached ONLY by the API key that opened the session. The detach-grace
// and fan-out retention windows are on by default, so without this check any event-scoped
// key that learns a session id could attach to another key's retained session and receive
// its replayed and live data. Ordinal comparison; null owner (session opened with no auth)
// matches only a null caller key.
if (!string.Equals(session.OwnerKeyId, callerKeyId, StringComparison.Ordinal))
{
throw new SessionManagerException(
SessionManagerErrorCode.PermissionDenied,
$"Session {request.SessionId} is owned by a different API key; event-stream attach is owner-scoped.");
}
// No `using` here — subscriber.Dispose() is called exactly once in the finally
// block below, which also disposes the reader. A `using` declaration would add a
// second Dispose on the same path and double-decrement the session subscriber count.
// The subscriber mode (single vs. multi) is derived inside AttachEventSubscriber from
// the session's own SessionEventStreaming.AllowMultipleEventSubscribers field — the
// same source the distributor uses — so the two cannot diverge.
//
// Reconnect/resume: when AfterWorkerSequence > 0 the client is resuming, so
// attach via the replay variant that atomically snapshots the replay ring AND registers
// the live subscriber under one lock. That single critical section is the crux of the
// no-gap/no-duplicate handoff: every replayed event has sequence <= LiveResumeSequence
// and every live event delivered below is filtered to sequence > LiveResumeSequence, so
// an event that was both replayed and (racing the registration) fanned into the live
// channel is dropped exactly once, while no newer event is skipped. See
// SessionEventDistributor.RegisterWithReplay for the full argument.
//
// AfterWorkerSequence == 0 (fresh stream, not a resume) keeps the original behavior:
// a plain attach, no replay, no sentinel, and the live filter watermark stays 0.
ulong afterWorkerSequence = request.AfterWorkerSequence;
IEventSubscriberLease subscriber;
IReadOnlyList<MxEvent> replayedEvents = [];
bool replayGap = false;
ulong oldestAvailableSequence = 0;
if (afterWorkerSequence > 0)
{
EventSubscriberReplayAttachment attachment = session.AttachEventSubscriberWithReplay(
options.Value.Sessions.MaxEventSubscribersPerSession,
afterWorkerSequence);
subscriber = attachment.Lease;
replayedEvents = attachment.ReplayedEvents;
replayGap = attachment.Gap;
oldestAvailableSequence = attachment.OldestAvailableSequence;
// The live filter resumes strictly after the last replayed sequence (or, when
// nothing was replayed, after the requested watermark). This is what makes the
// handoff free of duplicates: anything <= this watermark was already replayed.
afterWorkerSequence = attachment.LiveResumeSequence;
}
else
{
subscriber = session.AttachEventSubscriber(
options.Value.Sessions.MaxEventSubscribersPerSession);
}
IAsyncEnumerator<MxEvent> reader = subscriber.Reader
.ReadAllAsync(cancellationToken)
.GetAsyncEnumerator(cancellationToken);
// GWC-15: register this subscriber's channel as a live backlog source instead of
// reconciling the queue-depth gauge on every event. The gauge previously read the
// bounded channel's Count (which takes the channel's internal lock) and adjusted the
// metric under its own lock on every streamed event. Now the metric reads Count only
// when it is scraped (ObservableGauge callback) or projected (GetSnapshot), summing the
// live backlog across every registered subscriber — the same "buffered, not yet
// delivered" aggregate the per-event push reported, but with no per-event lock traffic.
// Disposing the registration in the finally removes this subscriber's contribution, so
// the gauge returns to the other subscribers' backlog (zero when none remain) on
// disconnect. CanCount guards a channel that ever cannot report Count (contributes 0).
IDisposable backlogRegistration = metrics.RegisterEventStreamBacklogSource(
() => subscriber.Reader.CanCount ? subscriber.Reader.Count : 0);
try
{
// Emit order for a resume: the ReplayGap sentinel FIRST (only when events were
// evicted), then the still-retained replay batch, then live. The sentinel is an
// explicit documented control signal (not a synthesized MXAccess event) and is
// delivered ONLY to this resuming subscriber — it is never fanned to other
// subscribers and never appears in DrainEventsReply (that path is untouched).
if (replayGap)
{
yield return CreateReplayGapSentinel(
request.SessionId,
request.AfterWorkerSequence,
oldestAvailableSequence);
}
foreach (MxEvent replayedEvent in replayedEvents)
{
// RegisterWithReplay already returns only events strictly newer than
// AfterWorkerSequence, so no per-item sequence guard is needed here.
// There is no per-event constraint filter on the event stream: events are
// fanned as-is by the distributor pump. The only dedup watermark is the
// LiveResumeSequence applied in the live loop below (to drop any event
// that was both replayed and raced into the live channel).
yield return replayedEvent;
}
while (true)
{
MxEvent mxEvent;
try
{
if (!await reader.MoveNextAsync().ConfigureAwait(false))
{
break;
}
mxEvent = reader.Current;
}
catch (WorkerClientException workerException)
{
// The distributor pump completes every subscriber channel with the source
// fault when the worker event stream terminates abnormally; that surfaces
// here. Mirror the original ProduceEventsAsync behavior: fault the
// session and record the metric, then propagate the terminal fault to the
// gRPC client.
session.MarkFaulted(workerException.Message);
metrics.Fault(WorkerClientErrorCode.WorkerFaulted.ToString());
throw;
}
// Per-RPC filter stays at the subscriber boundary: each request may resume
// from a different AfterWorkerSequence, so the shared pump fans raw events and
// this loop drops the ones at or below the caller's watermark.
if (mxEvent.WorkerSequence <= afterWorkerSequence)
{
continue;
}
// The queue-depth gauge is maintained lazily via the backlog registration above
// (GWC-15): the metric reads this subscriber's channel Count only when scraped,
// so there is no per-event gauge bookkeeping on this hot path.
yield return mxEvent;
}
}
finally
{
await reader.DisposeAsync().ConfigureAwait(false);
// Remove this subscriber's live backlog contribution before disposing the lease so
// the gauge stops counting a channel that is about to be completed; after this the
// gauge reflects only the remaining subscribers (zero when none remain).
backlogRegistration.Dispose();
subscriber.Dispose();
metrics.StreamDisconnected("Detached");
}
}
// Builds the single ReplayGap control sentinel emitted at the head of a resumed
// StreamEvents stream when the requested AfterWorkerSequence predates the oldest event
// still retained (events were evicted). Per the proto contract (MxEvent.replay_gap),
// the sentinel carries the session id and the populated ReplayGap, with family
// UNSPECIFIED, no body, and no per-item fields. It is a documented control signal — NOT a
// synthesized MXAccess event — so emitting it does not violate the no-synthesis rule.
private static MxEvent CreateReplayGapSentinel(
string sessionId,
ulong requestedAfterSequence,
ulong oldestAvailableSequence)
=> new()
{
SessionId = sessionId,
ReplayGap = new ReplayGap
{
RequestedAfterSequence = requestedAfterSequence,
OldestAvailableSequence = oldestAvailableSequence,
},
};
}