fix: merge count(*)::numeric across shards by murex971 · Pull Request #1936 · pgdogdev/pgdog

murex971 · 2026-11-02T15:36:31Z

Cross-shard count(*)::numeric failed with an invalid cast error: the parser treats a cast count as a count, but the merge only accepted integer and float partial counts.

Partial counts that come back as numeric are now converted to integers and summed, and the total is returned as numeric, the same as single-node Postgres.

codecov · 2026-11-02T15:39:21Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

levkk · 2026-11-02T16:14:03Z

Nice! We should add integration tests for this here: https://github.com/pgdogdev/pgdog/tree/bac0f8ea62e0ac73b871a0c06dd9c3224db81830/integration/sql

@jkaczman is also working on an acceptance test framework that compares the results returned by sharded and non-sharded pg, all the way to the binary protocol level, to ensure both are identical.

github-actions · 2026-11-02T19:27:04Z

Metal bench (arm64) passed. All 152 measures within bounds.

suite	throughput	cpu/txn	user cycles/txn	user instr/txn	kernel cycles/txn	kernel instr/txn	ctx/txn
pooler		20.02 (+0.2% ± 0.5%)	20,484 (+0.1% ± 0.4%)	35,285 (-0.0% ± 0.03%)	31,086 (+0.5% ± 0.8%)	51,876 (-0.0% ± 0.1%)
lb		20.59 (+0.2% ± 0.6%)	21,063 (-0.0% ± 0.3%)	36,197 (-0.0% ± 0.03%)	31,761 (+0.6% ± 1.0%)	53,060 (+0.0% ± 0.1%)
sharding-w2		38.3 (-0.6% ± 0.6%)	53,044 (-0.5% ± 1.0%)	102,575 (+0.0% ± 0.00%)	47,523 (-0.6% ± 1.1%)	54,286 (+0.0% ± 0.2%)	1.513 (+0.1% ± 0.2%)
sharding-w8	217,847 (+0.4% ± 0.4%)	33.66 (-0.3% ± 0.3%)	51,999 (+0.2% ± 0.4%)	57,890 (-0.0% ± 0.01%)	41,230 (-1.0% ± 0.3%)	42,923 (-0.0% ± 0.08%)	0.143 (-0.4% ± 1.7%)

Counters per txn

suite	frontend stall	l1i refill
pooler	8,454 (-0.0% ± 0.9%)	2,794 (+0.0% ± 0.3%)
lb	8,760 (+0.1% ± 0.6%)	2,853 (+0.3% ± 0.2%)
sharding-w2	24,593 (-1.6% ± 2.4%)	4,934 (-0.1% ± 1.5%)
sharding-w8	26,657 (+0.5% ± 0.9%)	4,793 (+0.4% ± 0.6%)

Microbench: all 135 measures within bounds.

lb flamegraph vs 3f7365a

Open interactive for hover, zoom and search.

Open interactive.

microbench flamegraph vs 3f7365a

Open interactive for hover, zoom and search.

Open interactive.

pooler flamegraph vs 3f7365a

Open interactive for hover, zoom and search.

Open interactive.

sharding-w2 flamegraph vs 3f7365a

Open interactive for hover, zoom and search.

Open interactive.

sharding-w8 flamegraph vs 3f7365a

Open interactive for hover, zoom and search.

Open interactive.

Logs

lb, microbench, pooler, sharding-w2, sharding-w8, verdict.json

Compared 1382fd8 with 3f7365a on m8gd.metal-24xl (spot).

github-actions · 2026-11-02T19:27:40Z

Metal bench (x86) passed. All 152 measures within bounds.

suite	throughput	cpu/txn	user cycles/txn	user instr/txn	kernel cycles/txn	kernel instr/txn	ctx/txn
pooler		35.27 (+0.3% ± 0.3%)	24,835 (+0.9% ± 0.3%)	36,684 (-0.0% ± 0.01%)	57,478 (-0.0% ± 0.4%)	38,353 (-0.1% ± 0.1%)
lb		36.35 (+0.5% ± 0.2%)	26,156 (+1.8% ± 0.6%)	37,632 (-0.0% ± 0.02%)	58,443 (+0.1% ± 0.2%)	39,063 (-0.1% ± 0.1%)
sharding-w2		57.85 (+0.7% ± 0.4%)	56,755 (+1.2% ± 0.6%)	106,017 (-0.0% ± 0.01%)	72,942 (+0.5% ± 1.0%)	43,127 (-0.0% ± 0.2%)	1.384 (-0.2% ± 0.1%)
sharding-w8	143,007 (-1.2% ± 0.4%)	49.89 (+0.3% ± 0.4%)	62,517 (+0.4% ± 0.2%)	61,421 (+0.0% ± 0.05%)	64,409 (+0.1% ± 0.8%)	34,254 (+0.1% ± 0.09%)	0.1567 (-0.2% ± 1.0%)

Counters per txn

suite	frontend stall	l1i refill
pooler	79,075 (+2.5% ± 0.5%)	3,847 (+2.4% ± 0.1%)
lb	83,270 (+3.8% ± 1.0%)	3,985 (+3.0% ± 0.2%)
sharding-w2	169,736 (+2.7% ± 0.9%)	6,293 (+1.6% ± 0.4%)
sharding-w8	138,259 (+2.8% ± 0.6%)	5,132 (+1.5% ± 0.2%)