Orca's full E2E run on main failed the same specs every night because they still expected old product behavior. My fix waited ten days while main fixed half of them; a maintainer rebased the rest, found the real cause of the failure I had written off as pre-existing, and merged it.
Argo CD v3.6.0-rc4 is the first release with the fix for a sharded controller that leaves a cluster without a shard. It is not backported to 3.5 or earlier. Here is how to check your version, spot the symptom, and work around it until you upgrade.
The sharding fix from the 57-hour incident merged into argoproj/argo-cd today. A timeline of the six-week review, the bots that review your PR before any human does, and the discovery that an out-of-date branch merges fine — no conflicts and green checks is what counts.
NServiceBus's delayed-messages poller retries without backoff when lock acquisition fails with a non-409 error. I filed the bug, wrote the two-file fix — and then read the CLA's indemnification clause and withdrew the PR.
A remote client could see my Windows workspace but none of its eleven tabs. Changing only the path separators in the workspace ID reproduced it — so I filed the bug and sent the fix.
A pmm-client sidecar on ProxySQL pods: secrets from Azure Key Vault via an init container, a stable registration name so rollouts don't litter the PMM inventory, and the four things that broke on the way (pmm-admin 3 flags, envsubst, a CA-less image, auto-sync).
Adding one cluster to a round-robin sharded Argo CD left two other clusters owned by no controller replica. 313 Applications froze at their last-known Synced/Healthy status for 57 hours, and CI kept reporting success. Here is how we found it, what the upstream cause turned out to be, and the one-line fix I sent to argo-cd.
After a five-year pause the blog has moved from WordPress.com to Hugo, hosted on Cloudflare Pages with the source on GitHub. All 140 posts, URLs, images and comments came along. Here is what changed and what is coming.