Redis — policy & pitfalls

Baseline Redis knowledge (strings, hashes, lists, sets, sorted sets, streams, pub/sub, TTL, MULTI/EXEC, basic replication) is assumed. This skill encodes the operational policy and the traps that keep showing up in production — the parts LLMs miss by default.

Setup Check (run first)

Before writing non-trivial Redis code:

1. Connection & topology — standalone, replicated (Sentinel), or Cluster? Cluster changes multi-key commands rules (hash tags required). Check INFO replication and CLUSTER INFO if unsure.
2. Redis version — KEEPTTL in SET requires 6.0+; COPY, GETEX, LPOS, ZADD GT/LT, XADD MINID require 6.2+; functions require 7.0+; sharded Pub/Sub (SPUBLISH/SSUBSCRIBE) requires 7.0+. INFO server | grep redis_version.

   Redis 8 note: Redis 8 unified Redis Community with Redis Stack — JSON, Search, TimeSeries, Bloom filters are now built-in, not add-ons. If running Redis 8+, these data types are available by default. All patterns in this skill apply equally to Redis 7.x and 8.x.

3. maxmemory + maxmemory-policy — CONFIG GET maxmemory-policy. Default noeviction → writes fail when full. Caches should be allkeys-lru / allkeys-lfu. Mixed (cache + durable data) → volatile-lru + TTL on cached keys only.
4. Persistence mode — RDB, AOF, both, or none. Affects restart semantics, fork memory, fsync latency. CONFIG GET save + CONFIG GET appendonly.
5. Client library — connection pooling configured? Sensible socket_timeout and retry_on_error? A lazy redis.Redis() with no pool is a production incident waiting to happen.

MUST DO

• Atomic lock acquisition: SET lock:<resource> <owner-token> NX EX <seconds> — one round trip, atomic. Use a random token for the owner so release is safe.
• Safe lock release via Lua (compare-and-delete) — not raw DEL:

  if redis.call("GET", KEYS[1]) == ARGV[1] then
      return redis.call("DEL", KEYS[1])
  end
  return 0
  

  Plain DEL would delete someone else's lock after your TTL expired.
• SCAN for iteration — never KEYS * on a production instance. SCAN is cursor-based, non-blocking, returns approximate matches per call.
• TTL on every cache key — set at write time (SET ... EX, SETEX, EXPIRE). A key without TTL lives forever and fills memory.
• Jitter TTL — base + random(0, base * 0.1) to avoid thundering herd when many keys expire simultaneously.
• Keyspace naming: <service>:<entity>:<id>[:<field>], colon-separated, ASCII, short but readable. Consistent prefixes enable observability and safe bulk operations.
• Pipelining for batches of independent commands — one round trip, N commands. Distinct from transactions.
• Idempotent writes — SET key value always wins; HSET k f v overwrites field. INCR is idempotent per-command, not per-retry — combine with a dedup key if at-most-once matters.

MUST NOT DO

• SETNX + EXPIRE as two commands. If the client dies between them, the key lives forever (lock leak). Always SET ... NX EX. The old SETNX command has no EX option — writing SETNX key val EX 30 is a syntax error.
• KEYS * / KEYS pattern in production. O(N) over the entire keyspace, blocks the server. Use SCAN with MATCH + COUNT.
• HGETALL on large hashes (thousands of fields) — single-blocking O(N). Use HSCAN, or redesign (split the hash, store separate keys).
• Multi-key commands across cluster slots without hash tags. MGET k1 k2 across shards returns CROSSSLOT. Force co-location with hash tags: {user:42}:profile, {user:42}:settings.
• EXPIRE on a non-existent key — silently returns 0. Check the path: some code sets TTL before insert.
• FLUSHDB / FLUSHALL / CONFIG RESETSTAT unprotected. Rename (rename-command FLUSHALL "" in redis.conf) or ACL-restrict these in production.
• Pub/Sub as a message queue. Pub/Sub is fire-and-forget — offline subscribers miss messages, no ACKs, no replay. Use Streams (XADD + consumer groups) for at-least-once delivery.
• Storing huge blobs (>100 KB per value). Redis is a RAM database; every replica / snapshot duplicates them. Use object storage + Redis for metadata.
• Treating Redis transactions as RDBMS transactions. MULTI/EXEC gives atomicity + isolation for the batch, but commands inside the block are queued without intermediate reads (use WATCH for optimistic locking). Errors in a command don't roll back others — syntactically invalid commands abort the batch; runtime errors let the rest execute.
• Blocking commands (BLPOP, BRPOP, XREAD BLOCK) on a shared connection — they tie up the whole connection for their timeout. Use a dedicated connection / pool for blockers.
• Caching with maxmemory-policy=noeviction — writes will fail with OOM when memory fills. Default of bare Redis; explicitly set allkeys-lru or allkeys-lfu for cache workloads.
• Using MULTI/EXEC in a cluster across slots — same CROSSSLOT rule; hash-tag co-locate or restructure.

Reference Guide

All extended patterns (data-structure trade-offs, lock correctness levels, cache stampede mitigation, cluster hash-tags, persistence & eviction details, observability) live in a single file: references/pitfalls.md. Load it when the task goes beyond the MUST rules above.

Output Format

When producing Redis code:

1. Short plan (1–3 bullets) — what keys / data structures / TTLs / atomicity guarantees are involved.
2. Commands or client code. Use language-agnostic Redis commands when the caller is unspecified.
3. For writes: state the eviction / TTL / persistence implications (e.g. "adds ~200 keys/min, TTL 1h, fits in allkeys-lru cache").
4. For multi-step atomic operations: state whether it's SET NX, MULTI/EXEC + WATCH, or Lua, and why.
5. Cluster-aware: if multi-key, show hash tags.

When reviewing Redis code: call out MUST-DO / MUST-NOT violations, especially missing TTL, non-atomic lock acquire/release, KEYS usage, and Pub/Sub used as a queue.