deckhouse/prompp v0.8.13

on GitHub

Features

1. Series carrying a label set but no samples are now rejected while the protobuf is parsed. The check already existed, but only fired later, when the series was added to the LSS during relabeling — a much heavier operation that takes locks and does extra bookkeeping for input that was going to be rejected anyway. read_timeseries_label_set / Protobuf::parse_timeseries now throw right away instead. The new opt-in PROMPP_FEATURES=enable_skip_no_samples_series flag silently skips such series instead of erroring, for clients known to send them legitimately (#506).
2. The WAL V2 segment format can now actually be turned on. The format (per-segment embedded ID, decoded by loadSegmentsV2) already existed, but createShardOnDisk always wrote V1 regardless: the enable_wal_writer_v2 PROMPP_FEATURES flag had nothing wired to it. Shard WAL headers and segments now go through a pair of process-global switches that EnableWalWriterV2() flips together, so the segment encoding and the header version tag stay consistent and replay picks the matching decoder (#510).

Fixes

1. Finalized chunks leaked when jemalloc arenas were enabled. FinalizedChunkList took no Reallocator and always allocated through plain malloc/free, bypassing the arena, so the finalized chunks of a DataStorage with arenas enabled were never freed when it was destroyed. It now takes a Reallocator, matching OutdatedChunk (#508).
2. Series still receiving samples over remote write could be pruned from the LSS as inactive. On the read-only transition path (input_transition_relabeling_only_read) a series found in the target LSS was not marked active, and the active-series bitset was not resized after an LSS copy, so a later prune before the fixed state could drop series that were in fact live. The path now marks found series active atomically, the copier resizes the bitset, and prune_hidden_series_before_fixed_state bounds itself by the next item index instead of the bitset size (#495).
3. LSS reallocations were lost when relabeling returned an error. The prompp_prometheus_per_goroutine_relabeler_* bridge functions computed build_deferred_indexes() and target_lss_has_reallocations only on the success path inside the try, and the appender skipped LSSResetSnapshot() whenever relabeling failed. Since allocations can happen before relabeling aborts, this left the head's allocated-memory tracking and LSS snapshot out of sync with an LSS that had actually grown. Both now run unconditionally (#504).
4. Dependency security updates. Bumped google.golang.org/grpc to v1.83.2, picking up the fix for CVE-2026-84445 — a request missing both the :authority and Host headers panics an xDS gRPC server, taking the whole process down (#502). The OpenTelemetry modules (otel, otel/sdk, otel/trace and the three otlptrace exporters) moved to v1.45.0 (#511, #512, #513, #514); the SDK builds its resource against semconv 1.43, so tracing.go was moved off semconv 1.41 — otherwise resource.New refuses to merge the two schema URLs and no tracer provider can be installed.

Other

1. inline __attribute__((always_inline)) replaced with the PROMPP_ALWAYS_INLINE macro throughout the C++ core (#496).
2. The web UI Jest suites are green again. @codemirror/state 6.5.x pulls in @marijn/find-cluster-break, whose package main is ESM-only, which aborted every codemirror-promql suite under Jest; it is now mapped to the shipped CommonJS build. The legacy react-app — whose enzyme 3.11 stack is incompatible with current cheerio — was dropped from make ui-test, mirroring upstream; it is still built and linted. A ui-tests CI job runs the module suites (#469).