fix: keep the bundled Redis accepting writes when a snapshot fails [ENG-1707]

[twk3]

Keeps the bundled Redis accepting writes when an RDB snapshot fails, and moves the app to the 2026-07-26-006 patch release. Today one failed snapshot makes it reject every write, which stops ingestion and crashloops the director.

Found on a self-hosted install (chart 0.8.0, app 2026-07-26-004, ENG-1707). Redis reached 24 GB with a full 50 GB volume, and every write returned:

MISCONF Redis is configured to save RDB snapshots, but it's currently unable to persist to disk. ... (stop-writes-on-bgsave-error option)

Why it happens

redis.commonConfiguration replaces Bitnami's default (appendonly yes, save "") with just the loadmodule line. So the bundled Redis runs its built-in save schedule (3600 1 300 100 60 10000) with stop-writes-on-bgsave-error yes. A snapshot that fails because the disk is full, or because its fork is killed for memory, turns into a full write outage until a later snapshot succeeds.

The change

• Add stop-writes-on-bgsave-error no. Snapshots stay on: Redis holds run:*:cache for in-progress runs and delayed BullMQ jobs (run timeouts, syncRunDB/syncInstance), and nothing recreates either after a restart. Refusing writes loses every incoming result for certain; continuing only loses data if Redis restarts before a snapshot succeeds.
• Document the headroom a snapshot needs (memory limit about 2× used_memory, volume at least 2× the snapshot file) and the rdb_last_bgsave_status check, in values.yaml and docs/eks/dependencies.md.

An external Redis is unaffected.

App version 2026-07-26-006

appVersion and currents.imageTag go from 2026-07-26-004 to 2026-07-26-006 (patches/release-2026-07-26-006 in currents). On top of 005's on-prem seat check fix, it adds:

• currents#4301: director passes async handler errors to next instead of exiting. This is the crashloop from the same outage.
• currents#4302: pm2 stops writing log files for the writer and director ECS. This is the writer's PM2_HOME emptyDir filling up.
• currents#4304: the dashboard no longer initializes Sentry in on-prem builds.
• ENG-1535 #3905: tr-steps:*:upload-state TTL 24h → 2h.
• ENG-1535 #3907: the term value is hashed in term_marker keys (TTL stays 24h).

Testing

• helm lint passes, and helm template --set redis.enabled=true renders the line into redis.conf.
• redis/redis-stack-server:7.2.0-v15 with /data on a 1 MB tmpfs, 30k keys, then BGSAVE ("No space left on device"):
  • Current config: SET, JSON.SET and the director's MULTI / HSETNX / EXPIRE / EXEC all fail with MISCONF, then EXECABORT, matching the customer's log.
  • This config: rdb_last_bgsave_status:err, and all three writes succeed.
• On a running install, redis-cli CONFIG SET stop-writes-on-bgsave-error no unblocks writes without a restart (checked on the current config).

Release Plan

Merge after the 2026-07-26-006 images are published by the on-premise-release pipeline. The chart version stays 0.8.0 for the release PR to bump.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}