26 articles tagged devops.
A nightly export ran green for six weeks while sending nothing, because the API key it used had an expiration date on it and a housekeeping cron deleted the record the day after it lapsed. The three ways a credential dies, why only one of them is loud, and the register, the gauge and the authenticated probe that turn a surprise outage into a calendar entry.
A security questionnaire asks for a CSP, you copy the nonce recipe everyone publishes, and your statically generated site quietly becomes a dynamically rendered one. Which directives cost nothing, why the nonce is the expensive part, and how we split a policy so the security lands where it actually matters.
Raising the body-size limit is the fix that keeps working right up until it doesn't. Why buffering user files through your application is a dead end, the presigned handshake that replaces it, the checks that still have to happen after the bytes land, and why the ERP should be handed a pointer instead of a payload.
The fix took forty minutes and a quarter of your users still hit the bug three weeks later, because a mobile release is a copy on somebody else's device. What over-the-air updates genuinely replace, the minimum-version gate you have to ship in v1 or never, why your API is permanently multi-version, and the autumn maintenance window nobody schedules.
Every self-hosted tool ships its own user table, and a leaver has to be removed from all of them. The identity provider we put in front of Grafana, Metabase, ArgoCD and the custom apps — which parts map cleanly, why Odoo is the awkward one, and what you owe the thing once everything depends on it.
One shared staging server is always broken, always blocked and always stale — and every fix for that is really a fix for it being shared. What it takes to give each pull request a real environment instead: the database copy, the seed nobody owns, the integrations that must fail closed, and the teardown that is the whole point.
A rename is the cheapest thing a database does and one of the most dangerous things a deploy does, because two versions of your application are always live at once. Expand, backfill, contract — the release boundary that makes it safe, the locks that bite anyway, and what changes when Odoo owns the schema.
A dark screen is a support ticket. A screen still showing last week's price is a lie in your store, and nothing in a normal monitoring stack will tell you about it. Content as a versioned bundle instead of a URL, day-parts that depend on a clock you don't control, expiry as a safety feature, and telemetry that reports what's actually on the glass.
Most maintenance contracts are a monthly invoice with a promise attached. The line items that make one real — patch cadence, backups you have restored, an SLA that says what a clock is and when it runs, and hours you can actually spend — plus the questions that expose a hollow one in five minutes, applied to ours as well as everyone else's.
Next.js will happily serve your JS bundles from a CDN with one config line. We stopped doing that. Hashed chunks belong to the build that emitted them; images outlive every deploy — and treating both as 'assets' is how a rollback turns into a cache-purge incident.
Swarm was the sensible choice in 2019; its ceiling is real in 2026. The strangler-pattern migration we run — same registry, same secrets model, one service at a time.
SaaS observability bills scale with pod count, not with incidents. The VictoriaMetrics + Grafana stack we run in every cluster we operate, and the alerting rule that matters more than the dashboard.
Not every deployment needs progressive delivery. Matching strategy to blast radius — what we roll, what we blue-green, and the one system that gets canaries.
Overprovision and you pay for idle cores; underprovision and the OOM killer visits at month-end close. How we set requests and limits from real usage percentiles instead of round numbers that sounded right.
You don't need a storage team to run stateful workloads on Kubernetes — you need to pick the right replication model. Where Longhorn's simplicity wins, where Ceph earns its complexity, and how we decide per cluster.
Every image we deploy is built by CI, signed by cosign and checked by Kyverno at admission. Why we run the policy in audit mode first, and the PR-preview exception that makes it survivable.
A backup nobody has restored is a hypothesis, not a plan. The monthly drill we run against Velero snapshots, Postgres dumps and filestore syncs — and what it's already caught before a client ever noticed.
Hundreds of images, charts and npm packages age every week across our clusters. Renovate turns that into reviewed pull requests, and CI gates plus a minimum release age turn most of those into merges nobody has to click.
Secrets in .env files end up in repos; sealed secrets end up unrotatable. The Infisical → External Secrets Operator → pod pipeline that keeps credentials out of both git and CI logs.
Four years running CloudNativePG under ERPs and web apps: declarative failover, WAL-archived backups and major-version upgrades as config, not 3 AM SSH sessions. The gotchas to plan for, and where we'd still reach for something else.
Every session starts from zero: the same stack briefing, the same corrections, the same tokens burned re-explaining decisions you made months ago. mem0 gives AI development a persistent memory — what it actually does, the numbers behind it, and why we self-host it on our own cluster.
A production Odoo container is not `docker run odoo`. Multi-process layouts, odoo.conf mounts, and why your Dockerfile should not install modules at runtime.
Build tools and git history do not belong in production images. How we use multi-stage Dockerfiles to ship lean Odoo containers under 500 MB.
Server logs tell you something broke after a user calls. Error tracking tells you first, with the full traceback and context. How we wire Odoo to Sentry or self-hosted GlitchTip with the OCA sentry module.
Remote turbo caching turns ten-minute builds into ten-second cache replays, and a self-hosted Woodpecker pipeline deploys every PR as its own preview instance. The setup behind this very website.
Malicious package versions are usually caught within days. pnpm's minimumReleaseAge makes you structurally immune to the window where they do damage.