The gateway emits a ReplayGap sentinel MxEvent at the head of a StreamEvents
stream resumed via after_worker_sequence when the requested cursor predates the
oldest retained event. Clients previously ignored it, silently mis-treating a
lossy resume as continuous. Each client now surfaces the sentinel as a distinct,
typed, non-terminal signal (never synthesized, never swallowed) so a consumer can
detect the gap and re-snapshot; resume contract is
after_worker_sequence = oldest_available_sequence - 1.
- .NET: MxEventStreamItem (IsReplayGap/ReplayGap/Event) via new StreamEventItemsAsync
+ AsStreamItemsAsync extension. Build clean, 87 passed.
- Go: EventResult.ReplayGap field + IsReplayGap(); ReplayGap type alias. build/vet/test clean.
- Rust: EventItem enum (Event/ReplayGap); EventStream now yields Result<EventItem, Error>;
CLI renders REPLAY_GAP line / replayGap JSON. fmt/check/test/clippy clean.
- Python: ReplayGap dataclass; stream_events yields pb.MxEvent | ReplayGap. 131 passed.
- Shared docs: ClientLibrariesDesign non-goals reframed (reconnect-replay protocol is
consumable; auto-reconnect stays a non-goal); CrossLanguageSmokeMatrix resume-gap note.
Java client is deferred to the windev batch (no local JRE); CLI-15 stays open until it lands.
Claude-Session: https://claude.ai/code/session_01DMXXvNuPekkkrTEyPNxEkW
Handles the new MxSparseArray wire type (proto field 19 on MxValue::Kind):
- value.rs: map SparseArrayValue to MxValueProjection::Unset (write-only; never emitted on read path)
- session.rs: add write_array_elements() that builds the sparse proto value and delegates to write()
- tests: three unit tests asserting proto shape, empty-elements case, and read-path Unset projection
- README: document write_array_elements default-fill semantics and bare-name [] normalisation
Client.Rust-022 Restored Error::MalformedReply for register / add_item /
add_item2 and the bulk-subscribe / read-bulk / write-bulk
dispatch arms so malformed-but-OK replies fail loudly
instead of returning Vec::new().
Client.Rust-023 Restored next_correlation_id and routed every CLI close /
stream-alarms / acknowledge-alarm / bench-read-bulk call
through it so each call carries a unique opaque token.
Client.Rust-024 Added round-trip tests for read_bulk / write_bulk /
write2_bulk / write_secured_bulk / write_secured2_bulk
plus stream_alarms and percentile_summary unit tests.
Client.Rust-025 RustClientDesign.md re-synced — new bulk SDK, alarms
surface, Error variants, CLI command list, and the
Windows stack workaround.
Client.Rust-026 Session::read_bulk now borrows a tag slice; bench-read-
bulk binds tags once outside the warm-up / steady-state
loops.
Client.Rust-027 .cargo/config.toml selector tightened to
cfg(all(windows, target_env = "msvc")) and comment
rewritten to match reality (release + debug ship the
8 MB reservation).
Client.Rust-028 run_batch removed the empty-line break; stdin EOF is
the only terminator.
Client.Rust-029 Re-applied Client.Rust-001 / 002 / 012 — added the
missing doc comments, renamed BulkReplyKind variants,
and replaced the clone-on-copy with a deref under lock
so cargo clippy -D warnings is clean.
All resolved at 2026-05-24; cargo fmt + check + clippy + test all green
(55 tests).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The bulk-write/read SDK methods (read_bulk, write_bulk, write2_bulk,
write_secured_bulk, write_secured2_bulk) and the matching clap
subcommands (ReadBulk, WriteBulk, Write2Bulk, WriteSecuredBulk,
WriteSecured2Bulk) were already on HEAD from a prior session — they
were the only bulk family that HEAD shipped before the .NET / Go /
Python / Java parallel ports. The one missing piece from the divergent
branch (commit f220908) was the BenchReadBulk benchmark harness.
mxgw-cli/src/main.rs adds:
- BenchReadBulk clap variant with flags --client-name,
--duration-seconds, --warmup-seconds, --bulk-size, --tag-start,
--tag-prefix, --tag-attribute, --timeout-ms, --json — defaults match
the .NET and Go benches.
- run_bench_read_bulk(): open-session → register → subscribe_bulk on
the synthesized TestMachine_NNN.TestChangingInt tags to populate the
worker value cache → warmup → steady-state loop with per-call
std::time::Instant capture → unsubscribe → close-session.
- BenchStats + LatencySummary structs and a percentile()
helper (nearest-rank with linear interpolation, matching the Go and
.NET implementations) so the cross-language JSON output is byte-for-
byte comparable. JSON schema: language / command / endpoint /
clientName / bulkSize / durationSeconds / warmupSeconds / durationMs
/ tags / totalCalls / successfulCalls / failedCalls /
totalReadResults / cachedReadResults / callsPerSecond /
latencyMs:{p50,p95,p99,max,mean}. scripts/bench-read-bulk.ps1 will
pick up the Rust line on its next run.
session.rs picks up minor tightening tied to the bulk SDK methods that
were already in the file (per-entry validation paths, BulkReplyKind
dispatch coverage) — no public-surface change.
Verification: cargo build --workspace clean (the 2 pre-existing
options.rs missing_docs warnings remain — out of scope); cargo test
--workspace 34/34 passing; cargo clippy --workspace --all-targets has
only the 3 pre-existing tolerated warnings (enum_variant_names on
BulkReplyKind, missing_docs on options.rs, clone_on_copy on
galaxy.rs:282). Manual smoke against live gateway on localhost:5120:
read-bulk on two TestMachine tags returned wasCached=true,
wasSuccessful=true; bench-read-bulk --duration-seconds 2
--warmup-seconds 1 --bulk-size 2 --json ran 363 calls / 181.35 calls
per second / p50=5.3 ms / p99=7.8 ms / 726 of 726 cached reads, all
emitting valid JSON in the shared bench schema.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Joseph Doherty