◆ Apache Cassandra

Optimize vector search performance

This is real work, not a feature someone invented — it comes from real job ads and real questions people asked. Below are four ready AI prompts: get it done, make it easy for the next person to say yes to, work out the right move when you are stuck, and stop it coming back.

4prompts
See the prompt↓All Apache Cassandra tasks

The same task, four prompts

today's deadline · the next reviewer · the stuck moment · the pattern
AExecute — do the immediate taskRun a full compaction on the 'customer_embeddings' table in the 'recommendations' keyspace on…+
Run a full compaction on the 'customer_embeddings' table in the 'recommendations' keyspace on the production cluster to reclaim disk space and improve read latency for the vector search queries.
when the reply comes backPush once: ask it to sharpen the weakest part, and to say what it assumed. Helpful?
BImprove — make it easier to acceptBefore we push the new product recommendation model to production, analyze the…+
Before we push the new product recommendation model to production, analyze the 'customer_embeddings' table's current vector search performance. Identify any partitions or queries that are consistently slow and suggest schema adjustments or indexing strategies to get latency under 50ms for 95% of requests.
when the reply comes backPush once: ask it to sharpen the weakest part, and to say what it assumed. Helpful?
CDecide — diagnose the stuck momentThe new product recommendation service is failing its 100ms SLA for vector search queries,…+
The new product recommendation service is failing to meet its 100ms SLA for vector search queries during peak hours, and the business is threatening to roll back the feature.
The new product recommendation service is failing its 100ms SLA for vector search queries, especially during peak traffic. We've scaled up the cluster, but it's not helping. I'm afraid we misconfigured the vector index or the partitioning strategy, and if we can't fix it by end of day Friday, the business will roll back the entire feature. What's the most likely bottleneck and the fastest way to diagnose it without causing further instability?
when the reply comes backPush once: ask it to sharpen the weakest part, and to say what it assumed. Helpful?
DBecome — change the patternWe keep having to scramble to optimize vector search performance after every major model…+
We keep having to scramble to optimize vector search performance after every major model deployment.
We keep having to scramble to optimize vector search performance after every major model deployment. This leads to late nights and missed deadlines. What habit should I change in our deployment process to proactively identify and address potential vector search bottlenecks before they impact production, rather than reacting to incidents?
when the reply comes backPush once: ask it to sharpen the weakest part, and to say what it assumed. Helpful?

Questions people actually ask

honest answers, no sign-up

Every task here was seen in the real world. Someone doing the job named it, a real job ad asked for it, or a lot of people asked about it online.

If nothing real showed a task, it is not on the page. That is the whole rule.

They are the same job approached four ways, because what you need depends on where you are.

Get it done today. Make it easy for the next person to say yes to. Work out the right move when you are stuck. Learn the pattern so the job stops coming back.

For most of these jobs it can carry the heavy thinking - draft it, sort it, check it, rehearse it with you.

It cannot sit in your chair, take the blame when a number is wrong, or notice what nobody wrote down. Let it do the first 80%. Keep the last 20% that is truly yours.

No. Copy any prompt and paste it into the AI you already use. No account, no score, no wall in the way.

Any of them. The prompts describe the work rather than naming a product, so they are not tied to one assistant.

That is also why they keep working when you switch.

Change it freely. Every prompt is a starting line, not a rule.

Put in your real numbers, your real names and your real deadline. The more you make it yours, the better the answer comes back.

The tasks come from real job ads, published job data and the questions people ask in public forums.

The steps come from Apache Cassandra's own documentation, with practitioner sources for the traps the manual does not mention.

Push once. Ask it to sharpen the weakest part and to say what it assumed.

Most wrong answers come from a missing detail rather than a bad prompt - tell it the thing it could not know.