redis replica expiration harness

This is the harness behind The Expired Keys Your Redis Replica Still Counts.

It runs a digest-pinned Redis 7.4 primary and a read replica and measures what a replica does with logically-expired keys, because replicas do not run their own expiration cycle, they hold a key until the primary propagates a DEL.

Three experiments:

• A. Ghost keyspace — set 1000 keys with a short TTL, let them expire, then compare the replica’s DBSIZE against how many of those keys still return a value on GET.
• B. Command behavior — one expired key, queried on the replica: which commands mask it and which still count it, plus the contrast that a read on the primary lazily deletes it while a read on the replica does not.
• C. DBSIZE drift — with the primary’s active-expire cycle on, sample DBSIZE on both nodes as a 5000-key batch expires, and report the peak replica-minus-primary gap.

These are laptop measurements demonstrating the mechanism, not production numbers. DEBUG SET-ACTIVE-EXPIRE 0 is used in A and B to make the timing deterministic (so the primary’s background sweeper isn’t racing the measurement); C runs with it on, the realistic default.

Run it

Docker with Compose v2, plus Python 3.9+.

cd benchmarks/redis-replica-expiry
docker compose up -d --wait          # primary on :6391, replica on :6392

python3 -m venv /tmp/rep-venv && source /tmp/rep-venv/bin/activate
pip install -r requirements.txt

python benchmark.py | tee results/summary.txt
docker compose down -v

The pre-3.2 contrast run

Replica read masking landed in Redis 3.2. To see the older, scarier behavior, where a GET on the replica returns the expired value itself, run the same script against a pinned Redis 3.0:

docker compose -f docker-compose.legacy.yml up -d --wait     # 3.0, same ports
RESULTS_DIR="$PWD/results/legacy" python benchmark.py | tee results/legacy/summary.txt
docker compose -f docker-compose.legacy.yml down -v

On 3.0 the replica returns owner for the expired lock and EXISTS says 1; on 7.4 both are masked. The docker-compose.legacy.yml image is amd64-only and runs under emulation on arm64 hosts (slow, but fine for this).

Results

• summary.txt — the captured console run used in the post (Redis 7.4.9).
• legacy/summary.txt — the same script on Redis 3.0.7 (unmasked reads).
• command_behavior.csv — the per-command table from experiment B.
• dbsize_timeline.csv — primary vs replica DBSIZE samples from experiment C.
• run_metadata.csv — Redis version, ports, and the headline numbers.

The mechanism (replicas wait for the primary’s DEL) is not version-specific. What changed across versions is how much the replica masks on read: nothing pre-3.2, and GET/EXISTS/TTL/SCAN by 7.4. DBSIZE counts the ghosts on every version.