Tell me about a change of yours that got stuck in review going back and forth.

During the credentials migration I had a change to the deploy pipeline's secret injection sit open for nine days across four rounds of comments. Two reviewers were arguing with each other in my thread about where injection belonged, and I was rewriting to match whichever of them had commented most recently. Round four had more comments than round one, which is when I admitted this was not converging. I stopped replying and sorted what was open. Of about sixty comments, three were genuine design disagreements and the rest were naming and ordering. So I posted a summary comment listing those three questions, marked everything else as either accepted or moved into a follow-up ticket, and asked for twenty minutes with both reviewers on a call. On the call the disagreement collapsed in roughly eight minutes: they held different assumptions about whether the pipeline or the runtime owned the lease, which no comment thread was going to surface. I wrote the decision back into the review so there was a record, and it merged the following morning. The delay had a price. Rotation lag on the services queued behind that change drifted from twelve days to twenty-seven while it sat. My rule now is that a third review round means I stop typing and ask for a call — it is a rule, not a judgement call I make when I am already frustrated.

Running that migration, I noticed the stalling was not one change, it was the pattern. Three of my team's four open changes were each on their fourth review round, and mine was one of them. We were four engineers and the reviewers sat in two other groups, so every thread was a slow negotiation with nobody empowered to end it. I took my own stuck change first, because I was not going to propose a rule I had not paid for myself. I closed it, reopened it in two pieces, and moved the design question that was actually blocking it out of the comments and into a one-page doc with two options. Then I proposed three things to both reviewing groups: every comment is labelled blocking or non-blocking by whoever writes it, non-blocking never holds a merge, and after a second unresolved round we default to a fifteen-minute call. I asked for it as a trial for the remainder of the migration rather than as a policy, which is why it was agreed instead of debated. Median time to merge went from six and a half days to a day and a half, and rotation lag across the thirty-one migrated services held under five. The labelling convention outlived the migration. The call rule mostly stopped being needed once comments were labelled, which told me the real problem had been ambiguity about what blocked.