The constraint is the question. It rules out the two approaches most candidates reach for, and the approach that survives has one failure mode that can take the customer's production database down.
Sync a customer's on-prem Postgres to your cloud at 10,000 updates/sec. You may not install anything on the DB server.
The constraint is the question. It rules out the two approaches most candidates reach for, and the approach that survives has one failure mode that can take the customer's production database down.
Updated Aug 2026 · Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.
Rejecting polling is table stakes; at this rate it is obviously wrong and saying so earns nothing. The two things that separate answers are noticing that logical decoding needs a parameter change requiring a restart, which is a negotiation with the customer's DBA rather than a technical step, and knowing what a replication slot does when your consumer stops. An abandoned slot retains write-ahead log on the primary indefinitely and fills its disk, which means a bug in your pipeline becomes an outage in their database. An engineer who has run this in production says that unprompted.
No comments yet — be the first to share your approach.
