DBLab Forge
Postgres experiments you can check
DBLab Forge runs controlled Postgres experiments on real cloud servers: same-server comparisons, predictions with falsifiers written before the first row when a study makes them, the full data, and a brief with every number traced to its source.
For teams with a Postgres decision that is expensive to get wrong: UUIDv4 or UUIDv7 before a table reaches billions of rows, which connection pooler before a traffic jump, what to run after pg_restore before switching traffic.
Closed beta, available to PostgresAI customers. The UUID study took about three days, one of them running experiments. Its brief is public.
What a Forge study looks like
A question about Postgres behavior becomes a study with a written design. Opinions and quick benchmarks are where it starts, not where it ends.
Predictions first
When a study makes predictions, each one is written down with a falsifier before the first row is inserted, and scored afterwards: held, not met, or falsified. The UUID study had fourteen.
Real servers, same-server comparisons
Experiments run on cloud servers, not laptops. Two servers of the same type are not the same server, so ratios between the variants are taken on one server.
Runs reported per server
Results are reported per server, so the spread between servers stays visible instead of disappearing into an average. The UUID study had seven runs, one run per cell; the dump/restore study, three restores and nine cold runs per step.
Full data and a brief
In the UUID study, counters were sampled every 10 seconds: WAL position, pg_stat_io, pg_stat_wal, buffer residency, wait events. The brief gives every number with its source table, the scored predictions, the corrections, and the prior work.
Review before trust
Before a ratio is published, the review step compares the servers it came from. In the UUID study, the first run showed int8 inserting 12% faster than UUIDv7; the review found that the int8 server had a faster disk, and on the same server the two keys insert at the same rate (how we caught it). That study also had sixteen independent reviews, eight in a hacker role and eight in a DBRE role, run on different AI models and posted on the study record with our responses.
Example studies
The four Forge studies published so far, newest first. Each was discussed on Postgres.FM; every number comes from the post or the public brief linked below it.
Postgres.FM 196 · November 13, 2026
What does a temp table per transaction cost?
23×
the TPS of a CTE doing the same work: 86,997 vs 3,783 for CREATE TEMP TABLE in each transaction. Postgres 18.6, 16 vCPU, no idle sessions.
- With 5,000 idle sessions and a 500 TPS target, temp-table DDL got 127 TPS in one run and 249 in another; 502 with no idle sessions.
- One create and drop sends 132 invalidation messages, and every backend must read them, idle ones too.
Postgres.FM 195 · November 6, 2026
Does an extension in its own image avoid the restart?
192–193 s
without new read-write connections: CloudNativePG restart with persistent sessions, for an extension image and for a fat image alike. Medians of 5.
- With switchover instead, the longest read-write outage was 6.4 s and 6.5 s.
- 18 PGDG extensions take 98 MiB as one image layer; the CloudNativePG pgvector image, 0.31 MiB.
Postgres.FM 194 · October 30, 2026
What does a random UUIDv4 primary key cost?
0.62–0.69×
the UUIDv4 insert rate vs an ordered key on the same server, at 100M rows. Postgres 18, seven runs.
- UUIDv4 wrote 1.5–1.9× the WAL per row of an ordered key, at 90M–100M rows.
- The UUIDv4 primary key was 33% larger than UUIDv7, at 100M rows.
Postgres.FM 193 · October 23, 2026
Dump/restore done: what to run before traffic?
32.6 s → 23 ms
one query, cold cache: right after pg_restore, then after VACUUM ANALYZE. 250M rows, 31 GiB heap, 16 GiB RAM, Postgres 18.6.
- After ANALYZE alone: 6.1 s, an Index Only Scan with 24,983 heap fetches.
- Medians of 9 cold runs; also tested on 19rc1.
Watch the discussions
Kirk Wolak and Nikolay Samokhvalov discuss each study on Postgres.FM; in episode 195, with Álvaro Hernández of StackGres.
Questions we have answered before
Forge is the third generation of the experiment harness PostgresAI has built since 2018. The first two generations ran experiments in our consulting work, many of them public in tests-and-benchmarks. The four studies above ran on Forge; the questions below come from the earlier generations. A few:
Locks and contention
Limits that arrive without warning
The cost of a setting or an extension
Every study reports its limits. Your tables, row sizes, load and hardware shift the ratios, so the answer to your question comes from an experiment designed for your case.
Closed beta for PostgresAI customers
Tell us the question you need answered, and we will say whether a Forge study fits it.