NeverBlink AI

Database reliability and performance, for AI with AI. Currently supporting #ClickHouse #OpenSearch #Elasticsearch.

[NeverBlink AI @neverblinkai]

NeverBlink now supports #ClickHouse: Cloud, self-hosted, and the Kubernetes operator. The worst ClickHouse failures never page you: merge backlogs, stalled TTLs, materialized views dropping data. We catch them while they're still warnings, with the fix attached.

https://t.co/Iut7EirwrN

[NeverBlink AI @neverblinkai]

https://t.co/4dRypR6RbZ

[NeverBlink AI @neverblinkai]

DORA is five metrics now (deployment rework rate joined the four), and the report renamed itself to State of AI-assisted Software Development. The research keeps landing on one point: the constraint isn't writing code faster. It's everything downstream of the merge.

[NeverBlink AI @neverblinkai]

pgvector keeps embeddings in the database you already run: no separate vector service, no ETL shuttling embeddings around, no second backup story. For plenty of teams that beats a dedicated vector DB on operational grounds alone. #PostgreSQL

[NeverBlink AI @neverblinkai]

A DROP TABLE stuck in an infinite retry loop, erroring every 5 seconds, invisible to every dashboard. Nobody graphs "operations that never finish." but an AI aware of it could find it.

It was found by a NeverBlink's agent reading error logs the way a DBA would. Monitoring only shows you what you thought to chart. Generic AI wouldn't know what to look for. NeverBlink did it.

[NeverBlink AI @neverblinkai]

7/ None of this slows you down. Investigations still drop from an hour to minutes. You're just not betting the database to get there.

[NeverBlink AI @neverblinkai]

1/ How to let an AI near your production database without losing sleep. The principles we've settled on while building NeverBlink:

[NeverBlink AI @neverblinkai]

6/ Treat confident urgency as a smell. Agentic debugging doesn't fail by hesitating, it fails by building an articulate story on one wrong premise. Guardrails matter most at the exact moment the agent is most sure of itself.

[NeverBlink AI @neverblinkai]

4/ Frontiers runs 11 self-hosted Elasticsearch clusters on it: 70% faster incident detection, 80% less time managing them.
Where does SRE ownership stop in your stack?
https://t.co/N6KOBJvKmb

[NeverBlink AI @neverblinkai]

1/ Ask an SRE team about their Kubernetes setup and you get a detailed answer.
Ask why the OpenSearch cluster went yellow last Tuesday and it gets vague.
The database is where site reliability engineering quietly stops.

[NeverBlink AI @neverblinkai]

3/ We built NeverBlink to be the AI SRE for your databases. It does the root-cause analysis on metrics, logs, queries and config, and tells you what to fix and why. Your team decides what gets applied. Engineers on call 24/7.

[NeverBlink AI @neverblinkai]

@arshaddotin if OpenSearch is crashing for you try connecting it to @neverblinkai to find the root cause - other than that look here for good discussion on paging: https://t.co/dIDdx1Avp9

[NeverBlink AI @neverblinkai]

@awais_khawar yes there's a lot going on now in terms of performance

[NeverBlink AI @neverblinkai]

@DirtyZero @HosseinGorji021 we support both OS and pgsql if you ever find yourself struggling with sizing or performance

[NeverBlink AI @neverblinkai]

PostgreSQL hit 55.6% in the Stack Overflow survey, its biggest single-year jump ever. And when AI agents spin up a database without being told which one, they reach for Postgres: over 80% of new databases on Neon were created by agents, not humans. #PostgreSQL

[NeverBlink AI @neverblinkai]

NeverBlink ranks database issues by severity and impact, attaches the root cause, and sends it to Slack, email or PagerDuty. Next to Datadog or Grafana, not instead of them. How many of last week's alerts led to anyone doing anything?

[NeverBlink AI @neverblinkai]

If your on-call engineer mutes a channel just to get through the night, your alerts stopped doing their job a while ago. On database clusters it happens the same way every time.

[NeverBlink AI @neverblinkai]

What helps: Alert on what users feel: latency, errors, rejected requests, indexing lag. Rank by impact. Yellow in dev is not rejected writes in prod. Put the why in the alert, not just 'heap 92%'. Delete alerts nobody acted on.