vs
Postgres vs MariaDB
2 dedicated vCPUs and 4 GB per database, 5 million orders unless noted. Shown: req/s, middle run.
Published results
Measured on dedicated cloud machines, each round from freshly loaded data. They hold for the setup named on each card.
vsvs
Postgres JSONB vs MongoDB
2 dedicated vCPUs and 4 GB per database, one million orders. Shown: highest req/s held with p95 under 50 ms.
Reads
New orders
Status updates
Shop mix, defaults
Shop mix, fastupdate = off
Shop mix, fastupdate = off, hosts swapped one run
New orders, fastupdate = off one run
Big orders (40 items): status updates one run
Big orders (40 items): reads one run
Merged PRs
What each merged pull request changed in its short run, base and PR in turns on one GitHub runner. ▲▼ is better, ▲▼ is worse, grey is no clear change. These numbers can't be compared with the published results above.
No merged PRs with numbers yet. The first one lands here when it merges.
Contribute
Think a database was set up badly, or the app uses it wrong? Change it and the pull request measures the difference.
Fork and change
Edit the app or the database settings in one comparison folder.
Open a pull request
The owner approves the run, so outside PRs don't start by themselves.
Short load test
App, k6 and each database on one runner, base, PR, PR, base, about a minute each.
Numbers on the PR
A comment shows req/s and p95 with arrows. Merged PRs appear on this page.
# run it on your machine first (needs Docker, Go and k6)
git clone https://github.com/<you>/db-wars && cd db-wars/postgres-vs-mongodb
docker compose -f ci/compose.yml up -d --wait postgres
cat ci/bench.conf # the seed size, the k6 options and the connection strings
Open to change
app/: queries, indexes, schema, connection poolci/compose.yml: database version and settings
Gets no numbers
.github/,k6/,ci/bench.conf,results/: the test stays the same for every PR- Settings that return a write before it reaches the disk, such as
fsync=off, or data in memory