babyredis

🧱 Redis-like commands, SQLite underneath, no server to run.

babyredis is a redis-py-shaped client backed by a single SQLite file (WAL mode). Think redka, but pure Python, in-process, and pip-installable with zero dependencies — sqlite3 ships with Python.

from babyredis import Redis

r = Redis("cache.db")          # or Redis(":memory:") for ephemeral
r.set("session:42", "ciara", ex=3600)
r.get("session:42")            # b'ciara'
r.incr("page:views")           # 1
r.ttl("session:42")            # 3600

Why?

• No server. No daemon to install, configure, monitor, or pay for. Your cache is a file next to your app — perfect for single-server deployments, CLIs, scripts, and side projects.
• Survives restarts. Unlike a dict (or fakeredis), data persists and can be shared across processes on the same machine.
• Familiar API. Method signatures mirror redis-py (set with ex/px/nx/xx/keepttl/get, ttl returning -2/-1, bytes responses with a decode_responses flag), so swapping a small app off Redis — or onto it later — is mostly an import change.

Honest trade-offs: local SQLite reads are fast (no TCP round-trip), but Redis will beat this on write throughput and tail latency, and there's no cross-machine networking. If you need pub/sub, clustering, or six-figure ops/sec, run real Redis. This is for everyone who doesn't.

Supported commands (v0.3)

Strings: set, get, getset, getdel, setex, psetex, setnx, append, strlen, mset, mget, incr/incrby, decr/decrby, incrbyfloat.

Hashes: hset (with mapping=), hget, hgetall, hdel, hexists, hkeys, hvals, hlen, hmget, hsetnx, hincrby, hincrbyfloat, hstrlen, hscan/hscan_iter.

Sets: sadd, srem, smembers, sismember, smismember, scard, spop, srandmember, smove, sinter, sunion, sdiff, sinterstore/sunionstore/sdiffstore, sscan/sscan_iter.

Sorted sets: zadd (nx/xx/gt/lt/ch/incr), zincrby, zscore, zmscore, zrem, zcard, zcount, zrange/zrevrange, zrangebyscore/zrevrangebyscore (with (-exclusive and ±inf bounds), zrank/zrevrank, zpopmin/zpopmax, zremrangebyrank/zremrangebyscore, zunionstore/zinterstore (with weights and SUM/MIN/MAX aggregation; plain sets participate at score 1.0), zscan/zscan_iter.

Lists: lpush, rpush, lpop/rpop (with count), llen, lrange, lindex, lset, lrem, ltrim, linsert, lmove, rpoplpush.

Keys & server: delete, exists, expire, pexpire, expireat, persist, ttl, pttl, keys (Redis glob patterns), scan/scan_iter, rename, renamenx, randomkey, type, dbsize, flushdb, ping.

Pipelines: pipeline() queues commands and runs them in a single SQLite transaction on execute() — unlike real Redis MULTI/EXEC, the batch is fully ACID: an error rolls back the whole batch.

with r.pipeline() as pipe:
    pipe.set("a", 1).incr("hits").rpush("log", "x")
    results = pipe.execute()   # [True, 1, 1]

Semantics follow Redis: operations against a key of the wrong type raise ResponseError (WRONGTYPE), removing a collection's last element removes the key, and TTLs apply per key across all types. Expired keys are invisible immediately and physically purged lazily, with a periodic sweep on writes.

Async

babyredis.aio mirrors redis.asyncio: same constructor, every command awaitable, *_iter helpers are async generators. Commands run in a worker thread so the event loop never blocks on SQLite I/O.

from babyredis.aio import Redis

async with Redis("cache.db") as r:
    await r.set("k", "v", ex=60)
    await r.get("k")
    async for key in r.scan_iter(match="user:*"):
        ...

Performance

Single-operation latency beats Redis-over-loopback ~6-10x (no TCP round trip) and fakeredis ~5x; a plain dict beats everything by ~40x; Redis wins back parity with pipelining and wins outright on concurrent write throughput. Full table, methodology, and honest caveats in docs/performance.md. Reproduce with python benchmarks/bench.py.

Concurrency

File-backed databases use one SQLite connection per thread: reads take no Python lock (WAL allows concurrent readers) and writes serialize through SQLite's own locking. Cross-process access to the same file is safe — writes are atomic via immediate transactions. :memory: databases use a single lock-guarded connection (SQLite memory databases can't be shared across connections).

Compatibility validation

tests/test_redka_compat.py ports the edge-case suite from redka's Go tests (score tie-breaking, list range/trim boundary permutations, store-variant overwrite semantics, atomic mset rollback, and more). Where redka deviates from Redis — silent wrong-type reads, keeping emptied keys, refusing cross-type rename, no negative zrange indices — babyredis follows Redis, and each divergence is noted in the test.

On top of that, tests/test_oracle.py runs property-based tests with fakeredis as the oracle: a Hypothesis state machine fires random command sequences at both clients and asserts identical results (or identical failures) after every step.

Testing

babyredis doubles as a Redis stand-in for tests. Opt into the bundled fixtures:

# conftest.py
pytest_plugins = ["babyredis.testing"]

def test_something(babyredis_client):
    babyredis_client.set("k", "v")

Install

pip install babyredis

Roadmap

• Type hints + py.typed

Development

pip install -e .
pip install pytest hypothesis fakeredis
pytest tests/