Just Ask
Ask me anything about software architecture, careers, PHP, Go and the craft of building software; I answer here, in the open, for everyone. (Page 4/6)
Whatever's on your mind, don't hold back. Questions reach me directly; I answer the good ones and publish them on this page. Your email is never published.
RPO 5min means continuous replication (cross-region streaming + PITR); RTO 30min means a warm standby ready to promote (IaC-provisioned, S3 CRR, Route 53 failover). Active-passive is the right fit for these numbers. But an untested plan is a guess — rehearse regularly.
Fix both layers: throttle/sample repetitive logs in the app ('×10,000' as one line) and back off retries; on the host use logrotate with size/count caps and write logs to a dedicated volume, not the root disk. Treat unbounded logging as a bug.
Per-IP limits lose to rotating proxies. Key a Redis-backed sliding-window on multiple dimensions — username + IP + subnet/ASN + UA + country; add account lockout and CAPTCHA, and push true volumetric DDoS to the edge (Cloudflare). Fail closed on the login path.
Decide in advance what's critical (browse, cart, checkout) and what's nice-to-have (recommendations, similar products, reviews); put the nice-to-haves behind feature flags and short timeouts. Under load a kill-switch turns them off or serves mock data. Don't improvise at 98% CPU — rehearse first.
Split one atomic transaction into a sequence of local transactions, each with a compensating action. An orchestrator drives the steps over a queue and runs compensations in reverse on failure. Make every step idempotent and publish via the outbox pattern.
Let a node be primary only if it holds a majority (quorum). Don't hand-roll failover; use Patroni + etcd/Consul (Raft), and have the old primary cut off from the quorum demote itself (fencing). Use an odd number of voting members.
Stop calling a failing dependency: count failures while CLOSED, fail fast to a fallback while OPEN, probe with trial calls while HALF-OPEN. But the real killer is the 30s timeout — pair the breaker with a short timeout, a bulkhead and a fallback.
Query-string busting is unreliable — some CDNs ignore the query in the cache key. The robust pattern is content-hashed filenames (`app.a8f9b2.js`): the name changes only when content changes, so you cache forever with `immutable`. The one thing you must never cache long is the HTML that points to those hashes.
`wildcard`/`regexp` scan at query time, don't use the inverted index, and burn CPU at 10M docs. Move the work to index time: build the prefixes when you index with `edge_ngram`, so autocomplete becomes a plain term lookup (milliseconds).
`file_get_contents`/`Storage::get` loads the whole file into RAM — fatal at 2-5 GB. Best of all, don't pipe the bytes through PHP: a presigned S3 URL or `X-Sendfile`. If you must, use `Storage::readStream` + `StreamedResponse` and emit chunk-by-chunk so memory stays flat at a few MB.