Walk me through how you onboard a new engineer onto your team.

My target for a joiner on the reconciliation service is that something of theirs is running in production inside two weeks, and that they can explain how a payment record moves through the pipeline without looking at a diagram. Day one is boring on purpose: environment, credentials, and a half-day where I walk the main flow with them on a whiteboard, including the two ways it usually breaks. I name one person as their default interrupt for the first month, and I say out loud that asking after fifteen minutes of being stuck is the expected behaviour, not a failure. The first task is chosen, not offered. I want something real that touches the main path and is cheap to get wrong. For the last person I hired in, that was tightening the retry backoff on the nightly reprocessing job. It went out on day nine, and it happened to take the overnight peak backlog from about 19,400 records to a little over 2,300, which was a nice accident but mostly it meant she had seen the whole path from change to graph. We paired twice a week for the first month, then once, then on request. The signal I watch is what they ask about in week three — if it is still setup and tooling rather than the domain, my front-loading failed, and that is exactly what made me add the whiteboard walkthrough after a previous joiner spent a fortnight guessing.

At group level I stopped trying to run onboarding and started making it survivable when the best mentor is busy. Across four squads the ramp quality depended entirely on who a joiner sat next to, and the spread showed: median time to a first production change ran from nine days on one squad to thirty-one on another. Three things fixed most of it. First, every squad had to publish a one-page path for a joiner's first two weeks, in their own words — I set the shape and refused to write the content, because a ramp nobody on the team authored is a ramp nobody maintains. Second, mentors were named and briefed rather than volunteered by proximity, and I ran an hour with each new mentor on what to hand over and what to answer with a question. Third, we built a sandbox exercise where a joiner drains a deliberately backed-up queue of about 12,400 messages and writes up what they found, which teaches our failure vocabulary in an afternoon. The median across the group came down to eleven days, and the spread between squads mattered more to me than the median. What I did not standardise was the first task, because the squads know their own safe surfaces better than I do, and centralising that would have made it ceremonial.