Benchmarks for XTDB 2, XTDB 1, Postgres, and Sqlite using queries and anonymized data from Yakread.

Setup

1. Download the zipfile chunks (about 9GB total)
2. Re-assemble with cat yakread_benchmark_input.zip.part-* > input.zip
3. Unzip and rename to ./import/
4. Start up a postgres docker container with ./postgres.sh
5. Run ./run.sh setup (this might take several hours and takes up about 150GB of disk space)

Running the benchmarks

1. Run ./run.sh benchmark. This prints tufte output to the console and writes it to results/*.edn
2. Run clj -M:run report to get a nice table comparing the p50 run times for each test across DBs. Reads from results/*.edn.

Notes

• The :user/timezone attribute didn't get serialized correctly by nippy, so the data ingest code removes it. It might already be removed from the anonymized input data.

• There is a half-baked datomic.clj file that I was using to set up benchmarks for Datomic Local. However data ingestion gradually got slower and slower, and after ~10 hours it had slowed to a crawl with still ~1 million entities left to go, so I gave up. Maybe an actual Datomic Pro instance would fare better.

• the xtdb2 and sqlite benchmarks remap document IDs: sqlite changes them from uuids to ints since sqlite doesn't have a native uuid type (though maybe blobs are fine?), while the xtdb2 benchmark keeps them as uuids but updates the prefixes to get better locality.

Document counts:

{"ad_clicks"  3341,
 "ad_credits" 184,
 "ads"        89
 "bulk_sends" 829,
 "digests"    23091,
 "feeds"      24468,
 "items"      17460648,
 "skips"      4494,
 "subs"       107597,
 "user_items" 236804,
 "users"      4835}