ParadeDB Benchmarks

Vector CI uses pdbench for recall calibration and k6 for latency and throughput:

SQL assets live under assets/<dataset>/, with engine-specific queries in assets/<dataset>/<engine>/. Shared CI actions build pdbench, start the cluster, install comparison extensions, report index builds, and run k6 with profiling. Dataset workflows own fixture checks, recall profiles, and result context. For MS MARCO 113M vector CI, see the MS MARCO pipeline.

RunsOn benchmarks default to Spot. For on-demand, add benchmark-nonspot before the PR benchmark trigger label, or select runner_type=on-demand in Run workflow (CLI: -f runner_type=on-demand). Stack Overflow comment-triggered runs also honor the label. Scheduled runs remain on Spot.

This is a basic, single-query latency, benchmarking suite for ParadeDB. It executes all of the queries belonging to a dataset, capturing both hot and cold measurements of their performance. After setup, queries are executed until a 3-run window shows <0.1% variance/mean (or 10 runs pass). Then --runs samples are taken for the hot measurements. The cold measurement is always recorded from the first query run.

Prerequisites

The benchmarking scripts require a Postgres database with pg_search installed. If you are building pg_search with cargo pgrx, make sure to build in --release mode. It also requires AWS credentials with access to the CI account’s paradedb-benchmark bucket to load data and restore snapshots.

pg_stat_statements

Query timing is done via pg_stat_statements, so you’ll need to configure it. The important settings are:

  • pg_stat_statements must be in shared_preload_libraries.
    • ALTER SYSTEM SET shared_preload_libraries = 'pg_search,pg_stat_statements';
  • After restarting Postgres, configure it:
  CREATE EXTENSION pg_stat_statements;
  ALTER SYSTEM SET pg_stat_statements.track_planning = on;
  ALTER SYSTEM SET pg_stat_statements.track = top;
  SELECT pg_reload_conf();

New sessions will use these settings after the configuration reload.

Usage

The benchmark subcommand runs benchmarks against an already-loaded heap. load-heap bulk-loads CSV data into Postgres, snapshot-heap and restore-heap manage pgBackRest heap snapshots, and sample/convert prepare source data. See DATASET_PREPARATION.md for the full dataset flow.

To reproduce a benchmark locally from existing CSV data, load the heap once and then run benchmark:

POSTGRES_URL="postgresql://localhost:28818/postgres"

cargo run --release -- load-heap \
  --url "${POSTGRES_URL}" \
  --dataset stackoverflow \
  --size 100k

cargo run --release -- benchmark \
  --url "${POSTGRES_URL}" \
  --dataset stackoverflow \
  --index bm25

Heap snapshots live at s3://paradedb-benchmark/snapshots/{dataset}/v1/{size}/pg_backrest/. To reuse that heap through pgBackRest, stop Postgres before snapshotting or restoring:

PGDATA="/home/runner/.pgrx/data-18"
BACKREST_ARGS=(
  --stanza bench
  --repo-bucket paradedb-benchmark
  --repo-path-prefix /snapshots
  --repo-region us-east-1
  --repo-endpoint s3.us-east-1.amazonaws.com
  --repo-s3-key-type shared
)

(cd ../pg_search && cargo pgrx stop pg18)
cargo run --release -- snapshot-heap \
  --dataset stackoverflow \
  --size 100k \
  --pgdata "${PGDATA}" \
  "${BACKREST_ARGS[@]}"
cargo run --release -- restore-heap \
  --dataset stackoverflow \
  --size 100k \
  --pgdata "${PGDATA}" \
  "${BACKREST_ARGS[@]}"
(cd ../pg_search && cargo pgrx start pg18)

For more options:

cargo run -- --help

Notable benchmark Options

  • --dataset defaults to “stackoverflow”
  • --index (required): Selects the index to build/benchmark, datasets/{dataset}/indexes/{index}.sql (e.g. bm25).
  • --clear-caches must be set to false if you’re running on a non-Linux system. (It defaults to true).
  • --skip-index: Including this skips index creation (and the after-create-index hook). Useful for iterating on queries against an already-indexed database.
  • --runs: How many warm samples to capture from each query. Defaults to 3.
  • --query-filter: Runs only the queries whose name ({query} or {query} - {variant}) matches this regex. The dirty (% visible) runs of a matching query also run.
  • --vacuum: Controls whether VACUUM FULL ANALYZE, followed by VACUUM ANALYZE to update the visibility map, runs before the queries. Defaults to true.

Notable Heap Options

  • load-heap --size: Selects sampled/{size}/csv under the data source.
  • load-heap --data-source: Overrides s3_base_path in datasets/{dataset}/config.toml.
  • snapshot-heap --pgdata / restore-heap --pgdata: Points pgBackRest at the stopped Postgres data directory. This can also be provided with PGDATA.
  • snapshot-heap --config / restore-heap --config: Uses an existing pgBackRest config instead of generating one.
  • snapshot-heap --repo-* / restore-heap --repo-*: Provides the S3 repository settings when generating a pgBackRest config.

Benchmarking a PR

The benchmark-stackoverflow label runs every Stack Overflow query on a PR. To run only some of them, add a conversation comment (not a review comment) with a regex. The workflow passes it to --query-filter:

@paradedb stackoverflow ^count_filter

With a size before the regex, the workflow runs only that dataset size. The 20m size takes most of a full run, so 100k or 1m gives a faster result:

@paradedb stackoverflow 1m ^count_filter

The commenter needs write access, and the PR branch must be in paradedb/paradedb. The workflow replies with a link to the run. Then it posts the results to the PR, one comment for each dataset size.

Datasets

Each benchmark run uses a single dataset located under datasets/$name. The heap must already be present — loaded by load-heap (which reads from the dataset’s data-source at the given --size) or restored from a snapshot.

The queries that are benchmarked for a dataset are located directly under datasets/$name/queries/, or under datasets/$name/queries/{index}/ when index variants need different SQL. Queries can be defined either as a nested directory (e.g. queries/{query}/ containing a README.md and one or more .sql files representing variants such as postgres.sql, hash_partitioned.sql, range_partitioned.sql) or as a standalone queries/{query}.sql file. Each SQL file contains a single query. When multiple variants are present in a query directory, they are benchmarked and charted together under the {query} - {variant} naming convention.

Dataset Directory Layout

datasets/{name}/:

  • config.toml
  • create_tables.sql
  • indexes/{index}.sql (one file per index variant, e.g. bm25; chosen with benchmark --index)
  • prewarm.sql
  • queries/{query}/*.sql or queries/{query}.sql (or under queries/{index}/)
  • after_create_index.sql (optional)

Preparing Datasets

For preparing and managing non-synthetic datasets (loading source data, sampling, and conversion), see DATASET_PREPARATION.md.