v2.1.3-ce — Parallel analytics, on by default

Released 2026-10-14 · GitHub release

v2.1.3 is a performance-tuning release: analytical scans now run in parallel across every core automatically, the GenAI inference scheduler auto-tunes to the machine it boots on, and the Linux binary ships stripped. No new SQL surface, no API changes — a drop-in upgrade from v2.1.2.

Why it matters

The analytical work that used to pin one core and make a report wait now spreads across all of them, with nothing to configure. Point lookups and transactional writes are unchanged — the new parallelism adds zero cost to the operational path.

Parallel analytical execution

SELECT DISTINCT, GROUP BY, and aggregate scans execute in parallel across the available CPU, automatically — no flags, no tuning.

Workload (189k-row DISTINCT)	v2.1.2	v2.1.3
single query latency	44 ms	3.8 ms (11.7×)
throughput under concurrency	121 q/s	851 q/s (7.0×)

Scaling tracks physical core count. Measured on an i5-10400F (6c/12t), single engine, independently validated result sets.

GenAI concurrency auto-tunes to the hardware

The inference scheduler now derives thread allocation, concurrency, and queue depth from the detected CPU and memory headroom instead of fixed defaults. Bursts of EMBED/GENERATE (bulk embedding, RAG ingestion, concurrent chat) are absorbed with deadline-bounded backpressure — requests wait on a bounded queue instead of being rejected with capacity errors. Single-stream GENERATE latency is unchanged-to-faster.

Under a 32-way EMBED burst, every embedding was written and verified valid (row count and vector cardinality asserted, not HTTP status) — zero rejections, zero silent drops.

Smaller, hardened binary

• The Linux build ships stripped — ~24% smaller (334 MB → 254 MB) and without the internal symbol map. A private debug sidecar is retained in CI for crash analysis and is not part of the public tarball.
• Crash-looping installs on resource-constrained machines now stop at a failed status (StartLimitIntervalSec/StartLimitBurst on the systemd units) instead of restarting forever and inflating install telemetry.

No OLTP cost

Point lookups, single-row UPDATE/DELETE, and transactional workloads perform identically to v2.1.2 (measured at parity). The speedups above apply to analytical queries (scans / aggregates / DISTINCT) and scale with physical core count — point/OLTP operations are at parity, not faster.

Upgrade

Drop-in swap from v2.1.x. Tarball installs: upgrade the whole extracted folder (the engine loads ONNX Runtime from lib/ beside the executable).

docker pull synapcores/community:v2.1.3-ce

Validation

Validator 246/246 on the published binary · recipe-cert pristine failing-set identical to v2.1.2 (zero new regressions) · point-op A/B at v2.1.2 parity (SELECT-1 ceiling 0.97×) · parallel DISTINCT 7.0× throughput / 11.7× single-query · GenAI autotune burst 0 rejections, all embeddings persisted and valid · binary confirmed stripped · zero SIGILL on non-AVX-512 (i5-10400F).