Commits on Sep 10, 2026

1. fix: resolve AWS deploy crash-loop and CDK/deploy issues found via re… …

   …al deployment
   
   Root-caused against a real cdk deploy, not from reading code alone:
   
   - backend-service-stack.ts: ECS healthCheckGracePeriodSeconds silently
     defaulted to 60s once an ALB target was attached, but the backend takes
     ~85s to finish booting (JVM + ADOT javaagent + Spring context). ECS killed
     every task mid-boot before it could ever answer, crash-looping a perfectly
     healthy application forever. Raised to 240s and made the ALB target-group
     interval/thresholds explicit instead of relying on defaults.
   - backend-service-stack.ts: MaxMetaspaceSize=192m was sized before the ADOT
     javaagent was added; the extra instrumentation classes exhausted it under
     real traffic (OutOfMemoryError: Metaspace), which took down the Redis/ES
     health indicators and made /actuator/health report DOWN. Raised to 384m.
   - backend-service-stack.ts: the container health check used CMD-SHELL, which
     ECS runs through /bin/sh -- dash in eclipse-temurin:25-jre, which does not
     implement /dev/tcp. The probe could never pass no matter how healthy the
     app was. Verified directly against the real image (old form fails even
     against a healthy server; fixed form passes/fails correctly). Switched to
     CMD + explicit bash.
   - UserServiceImpl.java: a new user's first page load fires several
     authenticated requests in parallel (feed + /users/me at minimum); all find
     no row and try to provision, and the losers trip users_cognito_sub_key.
     provision() unconditionally blamed a username collision instead, 500ing a
     legitimate sign-in. Now re-reads and returns the winner's row. Reproduced
     live against the deployed stack, fixed, and covered by a new test.
   - database-stack.ts: t4g -> t3 instance class (Graviton RDS capacity was
     tighter than Intel in eu-central-1) plus a configurable multiAz flag.
   - search-stack.ts: ECS health check retries capped at the real AWS limit
     (10, was 20) with startPeriod raised to preserve intent; Elasticsearch
     moved off PRIVATE_ISOLATED to PRIVATE_WITH_EGRESS so the Fargate task can
     reach ECR/CloudWatch Logs before the container starts.
   - compute-stack.ts, cache-stack.ts, event-streaming-stack.ts,
     rtmp-service-stack.ts, ws-sfu-stack.ts: security-group description strings
     fixed to the AWS-documented allowed charset (no em-dash, no bare ">").
   - infra/bin/infra.ts: threads the new dbMultiAz context flag through.
   - frontend/vite.config.ts: configurable VITE_CACHE_DIR so the dev server
     does not fight write permissions on a shared cache path.
   - .claude/launch.json: local dev-server launch config for the frontend.
   
   Verified: full infra CDK suite green (26 suites / 89 tests), backend
   UserServiceImplTest 4/4 including the new race test, and a real cdk deploy
   of all 29 stacks reached CREATE_COMPLETE with the backend ALB target group
   reporting healthy for the first time.
   
   Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

   

   voidstackloop and claude committed Sep 10, 2026
   Configuration menu

   • View commit details
   • Copy full SHA for 46af13b
   • Browse repository at this point

   Copy the full SHA

   46af13b View commit details

   Browse the repository at this point in the history