Chronos/DBWaitlist

Blog / Benchmarks

Where is Chronos DB slower than Postgres, DuckDB and Redis?

Analytics over large tables, some reports against Postgres, merges when a thousand agents commit at once, and raw vector-search speed. Our own numbers, with the reasons, and what Chronos is not.

Abhishek DG · 5 min read
  • Benchmarks
  • Postgres
  • Honest numbers

In four places: analytics over large tables, some reports against Postgres on the same cores, merges when a thousand agents commit at once, and raw vector-search speed against Redis. These are our own measurements, and we list them first in our benchmarks, above the results we win. If you are deciding whether Chronos fits, start with the losses.

What are the numbers?

AgainstWorkloadChronos DBThe other system
DuckDB 1.5.5Analytics reports over 200,000-row tables2.8–25.4 ms0.4–8.3 ms: 2.2–7× faster
Postgres 18Three of five reports over 200,000-row tables, both on 2 cores5.4, 40.5, 23.4 ms4.8, 29.8, 18.0 ms: 1.1–1.4× faster
Postgres 18CH-benCHmark Q3, one warehouse, both on 2 cores88 ms64 ms
Postgres 17Merge p50, 1,000 agents on disk, one thread per agent714–718 ms128 ms
Redis 8.10Vector search median, 76k and 1M real embeddings0.72–0.74 ms, 1.36 ms0.46 ms, 0.50 ms, at 75.0% and 74.2% recall against our 99.98% and 98.6%
LanceDB 0.39Loading 76,000 vectors of 1,536 dimensions4.7 s1.0 s

Every row has a section in BENCHMARKS.md with the method, the machine and the raw output.

Why is it slower at analytics?

Chronos stores rows, not columns. A report that reads two columns of a wide table still walks whole rows. DuckDB is built for exactly that work and is 2.2 to 7 times faster at it. If your agents mostly run large aggregations, Chronos is the wrong engine for that part; keep the warehouse you have.

Against Postgres the gap is smaller and closing. Release 0.5.3 and 0.5.4 made joins and scans faster: one analytical join, CH-benCHmark Q18, went from 2.6 times slower than Postgres 18 to 1.6 times faster on the same two cores. Q3 is still slower, 88 ms against 64.

Why are merges slow under extreme contention?

With 1,000 agents each on its own thread, all forking, writing 1,000 rows and merging on disk, the median merge takes 714 to 718 ms against Postgres’s 128 ms. Every merge into main waits its turn at the log. At 25 workers the median is 25 to 26 ms.

  • 1,000 threads1,000 agents fork, write 1,000 rows, mergemerge p50 714–718 ms
  • 25 workersthe same 1,000 agentsmerge p50 25–26 ms
  • 64 workersthe same, rows written per second713k–716k rows/s; Dolt 28.3k, Postgres template databases 9.8k
The same workload, two ways of running it. Throughput stays high because agents work in parallel in their own worlds; it is the wait at the merge that grows.

The cause is known and the fix is being built. Until then, run a thousand agents through a pool of workers, not a thousand threads.

Why does Redis answer vector searches faster?

Because it finds fewer of the right answers. Redis returns in about half a millisecond with 74 to 75 percent of the true top ten. Chronos takes 0.72 to 1.36 ms and finds 98.6 to 99.98 percent. Speed and recall trade against each other, so we print recall beside every speed. If three in four is good enough for your search, Redis is faster.

What isn’t Chronos at all?

  • It is not Postgres. It runs much of Postgres 16’s SQL, not all of it: about 89 percent of the expected lines of Postgres’s own regression suite. Postgres compatibility lists what is missing.
  • It is a beta. One engineer writes the engine, and the storage engine has not had an outside review.
  • It is one process per database folder. No hot standby, no automatic failover.

How should I read these?

What Chronos is built for is the other half of the page: forking a branch in about 2 microseconds at any size, thousands of agents working in isolation, and merging their work back under rules. If that is your problem, the losses above are the price. If it is not, they are the reason to use something else.

Questions

Is Chronos DB faster than Postgres?

Not across the board. On the same two cores Postgres 18 is 1.1 to 1.4 times faster on three of five reports over 200,000-row tables, and faster on CH-benCHmark Q3. Chronos is much faster at what it is built for: forking a branch of the whole database and letting many agents write in isolation.

Is Chronos DB good for analytics?

No. It stores rows, not columns, and DuckDB is 2.2 to 7 times faster on analytics reports over 200,000-row tables. Keep a warehouse for large aggregations.

Why does Chronos publish the benchmarks it loses?

Because a result you cannot check against its failures tells you little. The losses are listed first in BENCHMARKS.md, each with its method and raw output, so you can judge fit before you try it.

Written by

Abhishek DG

Co-founder & CEO, 2090 Labs

Abhishek builds Chronos in Bangalore: the database where AI agents work in their own copy of your data. Before it he built TailsMart and ACRED India.

More about Abhishek

Keep reading

Where is Chronos DB slower than Postgres, DuckDB and Redis? · Chronos