Most Pulsar P1s trace back to a sizing choice made at build time. These are the questions we put to a team in the first week of a support contract, with the reason each one matters.
What are your ensemble, write quorum and ack quorum? A common starting point is an ensemble of 3, a write quorum of 3 and an ack quorum of 2, which tolerates one slow or failed bookie without stalling writes. A larger ensemble spreads load across more bookies but widens the blast radius of a slow disk. Where a team runs an ack quorum equal to the write quorum, every write waits for the slowest bookie, and one degraded disk becomes a cluster-wide latency problem. We check the setting against the number of bookies, the failure domains (rack or zone awareness) and the latency budget.
Where does the journal live, and what shares it? The journal is a sequential, fsync-heavy write path. It wants its own fast device. Teams that put journal and ledger directories on the same volume see compaction and ledger flushes show up as publish latency. We measure fsync latency on the journal device before the first production topic.
How many bundles does each namespace have, and how many topics are hot? Topics map to bundles by hash. A namespace created with the default bundle count and a few very busy topics can put most traffic on one broker. Splitting bundles early costs little; splitting under load causes the reconnect blips we see in problem 4. We size bundles from the expected topic count and the busiest topics' rates.
Which topics need compaction, and who reads them? Compacted topics are cheap until a compaction falls behind and a consumer reading the compacted view sees stale data. We check compaction thresholds and run state per topic.
What do Pulsar Functions and connectors share with the brokers? Functions running inside the broker process share its memory and CPU. A leaking function becomes a broker incident. We prefer separate function workers for production and check their resource limits.
What happens to a subscription when its consumers go away? Durable subscriptions keep their cursor and backlog forever unless a TTL or expiry policy removes them. A forgotten subscription from a retired service is the most common source of a backlog that grows for weeks and then blocks producers. We list subscriptions with backlog and no active consumer as part of every review.