Deep Engineering

Interview preparation

SRE

18 lessons in this track

  1. 1Signals and PID 1Signals, zombies and PID 1: why a container ignores SIGTERM25 MIN
  2. 2cgroups and OOMcgroups and the OOM killer: who gets killed, why the log is empty, and what 137 really meansClearer after: Signals and PID 125 MIN
  3. 3Descriptor limitsToo many open files: whose limit it is, why it is not shared, and what counts as a descriptorClearer after: Signals and PID 120 MIN
  4. 4Page cache and fsyncWhy "written" and "will not be lost" are different claimsClearer after: Signals and PID 125 MIN
  5. 5The accept queueThe connection is established and the server knows nothing about itClearer after: Descriptor limits20 MIN
  6. 6Timeouts and deadlinesThe client left, the work carried onClearer after: The accept queue25 MIN
  7. 7Retries and jitterWhy a service falls over a second time at the moment it recoversClearer after: Timeouts and deadlines20 MIN
  8. 8Graceful shutdownWhere the errors on every rollout come fromClearer after: Signals and PID 1 · The accept queue20 MIN
  9. 9Queues and tailsWhy 19 % more load makes latency four times worseClearer after: The accept queue25 MIN
  10. 10SLOs and budgetsSLOs and error budgets: why "99.9 %" without a window means nothingClearer after: Queues and tails25 MIN
  11. 11Health checksThree different questions called by one nameClearer after: Graceful shutdown · SLOs and budgets20 MIN
  12. 12Rollouts and rollbacksWhy a canary does not help without the right detectorClearer after: Graceful shutdown · Health checks25 MIN
  13. 13Name resolutionThe application calls getaddrinfo, not DNSClearer after: Timeouts and deadlines30 MIN
  14. 14TLS handshakeThe TLS handshake: what the first connection pays forClearer after: Name resolution25 MIN
  15. 15TIME_WAIT and portsTIME_WAIT and ephemeral ports: the client hits the wall firstClearer after: TLS handshake25 MIN
  16. 16Load balancingWhy two random choices beat round-robinClearer after: Queues and tails25 MIN
  17. 17Consistent hashingHow many keys move when the membership changesClearer after: Load balancing25 MIN
  18. 18IdempotencyExactly once is about the effect, not the deliveryClearer after: Retries and jitter25 MIN

Final check: the SRE track

One question per lesson. Not a recap but the places where the mechanism answers differently from experience: a 137 with not one line in the log, a peak equal to the limit, a connection to a server that never called accept, a second nameserver that doubles the worst case, a TLS resumption that saves no round trip, and a canary that lost to a blunt full switch.

Final check1 / 18

An application is PID 1 in a container and installs no SIGTERM handler. The orchestrator sends SIGTERM with a 30-second grace period. What happens?