No universal benchmark numbers are claimed. Repository benches and perf probes are development measurements under their own conditions — cite conditions when quoting them.
Levers you control#
- Index the predicate: equality/range/order/time-window columns first; expression indexes for hot JSON paths.
- Shape the query: selective
WHEREbefore joins;LIMITearly; prefer keyset over largeOFFSET;EXPLAIN ANALYZEactuals over estimates. - Batch writes: multi-row
VALUES,INSERT…SELECT, short transactions; group commit already coalesces concurrent commits. - Mind checkpoints: 64 MiB / 16 segments / 15 min defaults bound replay; raising thresholds defers checkpoint cost but lengthens recovery.
- Measure:
./target/debug/plomid benchmark run --iterations N --jsonwrites<data_dir>/benchmarks/<run>.json(avg/p50/p95/p99/throughput forSELECT 1).
./target/debug/plomid metrics reports probe latency plus live server counters from plomid_stat_counters() (queries, errors, connections, slow queries, maintenance passes, checkpoints, stale tables, TLS connections/failures).
Hardware tiers and SIMD#
At startup the server detects CPUs and memory and sizes itself: worker fan-out, memory budgets, batch sizes, and background frequency follow the tier. Container CPU/memory limits only shrink the sizing, never grow it.
| Tier | When | Buffer pool pages |
|---|---|---|
edge |
small devices | 256 |
laptop |
default developer machines | 1024 |
server |
≥16 CPUs and ≥32 GiB RAM | 4096 |
Override with --profile {edge|laptop|server|auto} (PLOMID_PROFILE); auto detects. Typos fail closed at startup. The active tier is reported in the startup log. Distance and analytics kernels dispatch to the widest CPU instructions the host supports (AVX-512 / AVX-2 / NEON / portable scalar) with identical results — there is no GPU execution path in v0.1.0.
Spill and memory budgets#
Large sorts and aggregations spill to disk instead of failing: past 250,000 rows, ORDER BY switches to a k-way merge and GROUP BY to hash-partitioned pre-aggregation, bounded by a 1 GiB per-operation temp limit (not user-configurable). Memory budgets are tiered internally (global/OLAP/maintenance/query/connection/worker/result-buffer/temp pools); over-budget operations fail with descriptive errors (memory budget exceeded, spill storage exhausted) rather than growing without bound. Large results stream with backpressure instead of buffering whole — there are no SQL cursors (DECLARE/FETCH unsupported), and none are needed for this.
HTAP governor#
Under memory pressure the engine protects short reads/writes first: an admission governor admits, queues, or rejects analytical work across 70/85/95% pressure bands while OLTP keeps a reserved lane, and background maintenance (compaction) pauses instead of competing with foreground queries. Watch plomid_htap_stats() (admission counts, pressure, spill bytes) and plomid_stat_contention() (mutex/gate/barrier waits) to see it act. The HTAP Gauntlet measures the resulting penalty — note its headline penalty is a p99-latency ratio (combined vs alone), not a throughput ratio.