A query planner's bad row-count estimate can be more expensive than a slow disk.
It may select the wrong join order or algorithm.
Correct metadata shape the right execution plan
A join can be cheap on one machine and expensive across a cluster.
If matching rows live on different workers, the engine may shuffle data over the network.
Join keys and partition layout can matter as much as SQL syntax.
A delete in an Cassandra database writes a tombstone rather than removing bytes immediately.
Reads and compaction must account for it.
Deleting more data can temporarily increase write work before space is reclaimed.
In object storage, 10,000 tiny files aren't equivalent to one large file—even if total bytes match.
Listing, metadata, request overhead, and task scheduling all add up.
File count is a performance dimension of its own.
Time-series partitioning by month makes retention convenient.
It can also make a query spanning month boundaries touch more partitions and files.
A partition boundary is an operational choice and a query-planning choice at once.
Columnar formats win when queries read a few columns from many rows.
But fetching one complete record can favor row-oriented access.
The storage layout should follow the query shape—not whichever format sounds more modern.
Kafka preserves order within a partition, not across an entire topic.
Choose the key poorly and related events may land in different partitions.
Choose it too narrowly and one partition becomes hot.
Partition keys encode both semantics and capacity.
Partitioning by hash distributes keys.
It doesn't necessarily distribute work.
If one key receives 40% of traffic, hashing still sends that key's requests to the same partition.
Balanced storage isn't the same as balanced load.
Interviewer: “Tell me about a time you showed leadership.”
Me: “Production went down. Nobody knew what to do.”
“What did you do?”
“Started a call.”
“And then?”
“Asked everyone for updates.”
“Welcome aboard. We’ll put you in management.”