Redis in production

Redis 7.2 plays four distinct roles in the Avo stack: regime state cache, signal pub/sub transport, job queue for the outbound email system, and per-tier API rate limiting.

The regime cache holds 209,033 keys in the argus:regime:* namespace. Each key stores a hash of regime scores across 5 regime types and 8 timeframes for a given symbol. Keys are written by the argus-regime Rust binary after every computation cycle. A significant bug was discovered in April 2026: the EXPIRE call was missing after the HSET write, meaning keys accumulated indefinitely. If the regime pipeline stopped, stale data would silently persist and be served to Apex as current. The fix set a 24-hour TTL on all 209K keys via a Redis SCAN loop and added the EXPIRE call to the Rust writer. The 24h window was chosen because regime state changes on a daily cycle; any key older than 24h is stale by definition.

Signal pub/sub uses Redis Streams with consumer groups. The Argus signal pipeline publishes computed signals to a Redis stream; Apex reads from the stream with at-least-once delivery guarantees. If Apex crashes mid-read, it resumes from its last acknowledged message ID. The stream is trimmed to 7 days of history with MAXLEN.

Example workflow: setting up a Redis Stream for signal delivery between two services. 1. The producer (argus-signals Rust binary) calls XADD signals:live * signal_id {uuid} symbol AAPL direction long confidence 0.82 after each computation cycle. 2. The consumer (apex-executor Rust binary) calls XREADGROUP GROUP apex-group apex-consumer-1 COUNT 10 BLOCK 1000 STREAMS signals:live > on startup. 3. On successful execution, the consumer calls XACK signals:live apex-group {message_id} to confirm delivery. 4. If the consumer crashes, unacknowledged messages stay in the Pending Entries List. On restart, the consumer calls XAUTOCLAIM to reclaim messages idle for more than 30 seconds. 5. Add a XTRIM signals:live MAXLEN 50000 call in the producer after every XADD to cap stream size at approximately 7 days of history. 6. Monitor stream lag with XLEN signals:live and XPENDING signals:live apex-group. Alert if pending count exceeds 1,000 (consumer is falling behind).

Job queues for the outbound email system (alien/cadence) use Redis Lists. Each outbound touch (email, DM, retarget post) is pushed onto a queue; a Rust worker pops and executes with configurable rate limits per inbox.

API rate limiting uses Redis INCR with EXPIRE in a sliding-window pattern. Each Clerk user tier (Free, Pro, Institutional) gets a different per-minute limit. The implementation is in the Next.js API layer, using ioredis from Node.js.

The caching layer for the Avo marketing site API routes was added in April 2026 after a performance audit: the /api/macro route dropped from 1,476ms to 465ms with a Redis cache; the /api/intelligence/regime route dropped from 1,048ms to 79-90ms with a collection cache.