Benchmarks

Compare pg_redis storage modes against Valkey with redis-benchmark.

make bench runs mise’s redis-benchmark with identical flags against pg_redis () and Valkey ().

Default pairing

The default comparison is the closest durable pair:

pg_redis — logged tables with synchronous commit
Valkey — AOF with appendfsync always

That is the closest available crash-durability comparison; the implementations still have different transaction and batching semantics.

make bench                 # current PG_STORAGE_MODE (default unlogged) vs Valkey
make bench-matrix          # memory,unlogged,logged-async,logged × Valkey modes
make bench-publish         # refresh README + bench/latest.{csv,md}
PG_STORAGE_MODE=memory make bench-pg

Matrix modes

make bench-matrix runs three repetitions per configuration and reports median, minimum, and maximum throughput:

Target mode	Persistence behavior
`pg_redis/memory`	In-gateway map (max throughput); optional snapshots off during benches
`pg_redis/unlogged`	No table WAL; survives clean restart, truncates after crash
`pg_redis/logged-async`	WAL-logged, but commit returns before WAL is flushed
`pg_redis/logged`	Normal WAL-logged tables; commit waits for durable WAL
`valkey/none`	No RDB or AOF; process data is volatile
`valkey/rdb`	Valkey’s default periodic snapshots; AOF off
`valkey/aof-everysec`	AOF with an approximately one-second loss window
`valkey/aof-always`	AOF fsync for each command/batch before reply

Default tests are ping,set,get (-t ping exercises both inline PING\r\n and RESP PING_MBULK).

Tuning

Variable	Role
`BENCH_REQUESTS`	Requests per test
`BENCH_CLIENTS`	Concurrent clients
`BENCH_PIPELINE`	Pipeline depth
`BENCH_MATRIX_REPETITIONS`	Repetitions for matrix medians
`PG_STORAGE_MODE`	pg_redis mode for single benches
`BASELINE_MODE`	Valkey mode for single benches

PG_STORAGE_MODE=memory BASELINE_MODE=none make bench
PG_STORAGE_MODE=unlogged BASELINE_MODE=none make bench
BENCH_REQUESTS=50000 BENCH_MATRIX_REPETITIONS=5 make bench-matrix
BENCH_REQUESTS=50000 BENCH_CLIENTS=20 BENCH_PIPELINE=16 make bench
./scripts/run_bench.sh --pg-only          # only :6379
./scripts/run_bench.sh --baseline-only    # only :6380

Published results

make bench-publish runs the matrix and refreshes:

the generated benchmark block in README.md
tracked bench/latest.md
tracked bench/latest.csv

Keep those three generated files in the same commit.

Single comparisons write CSV + a table under bench/results/<timestamp>/. Matrices write all repetitions, matrix.csv, and comparison.txt under bench/results/matrix-<timestamp>-<pid>/.

See the latest committed summary in bench/latest.md.