ML circle thread #20: what we log per prediction for audits
ML circle thread #20: what we log per prediction for audits
Ground this in something you measured or shipped — theory welcome, receipts preferred.
Thread index 20 — add your angle.
15 replies
The mentor who said 'measure twice, cut once' applied to migrations too literally. In ML practice (thread 20): The mentor who shared a bad estimate story made me kinder when others miss. The architecture review that asked about cold start SLOs changed our packaging strategy honestly.
The mentor who reviewed my incident timeline taught me to log timestamps always. In ML practice (thread 20): The incident commander who time-boxed debates saved minutes that mattered. We should have asked finance about chargeback patterns before promising SLAs in sales.
The quiet deletion of duplicate monitors reduced alert fatigue measurably. In ML practice (thread 20): The design that considered translation for circle descriptions improved international join rates measurably without noisy defaults always. We should have named a communications approver for incidents before execs posted early tweets.
We stopped confusing 'innovation' with 'new dependencies' when reviewing proposals. In ML practice (thread 20): The architecture review that asked about thread export for compliance changed retention policy design honestly before launch. The quiet win was aligning on a single customer health score definition across CS and product.
We learned that customers notice faster search more than a slightly prettier button. In ML practice (thread 20): The mentor who said 'show me the cohort retention table' ended faith-based growth debates again. We should have asked support what they hear before prioritising the roadmap.
The mentor who said 'write the customer apology draft before launch' improved incident comms. In ML practice (thread 20): We should have invested in local dev parity earlier; 'works on my machine' was expensive. We learned that writing down assumptions turns arguments into experiments faster.
The smallest type annotation prevented a class of null surprises — types as docs. In ML practice (thread 20): The design that considered one-handed mobile use caught a real thumb reach issue in testing. Junior devs spotted the smell first; seniors were too used to the workaround.
We should have deleted unused 'shadow' circles created during QA seed scripts — they polluted discovery until we purged them measurably. In ML practice (thread 20): We should have deleted unused API keys embedded in old Postman collections — security scanners still find them years later quarterly always. We learned that customers trust companies that publish post-incident learnings without corporate jargon.
We learned that writing postmortems for near misses prevents the big miss later. In ML practice (thread 20): The mentor who said 'prove it in staging' shortened debates with confident opinions. We learned that customers trust companies that admit mistakes in public status updates quickly.
We learned that transparent incident timelines reduce conspiracy theories internally too. In ML practice (thread 20): The mentor who reviewed my incident timeline taught me to log timestamps always. The quiet win was documenting which Redis cluster is cache vs session vs rate limit — three names, three owners.
The mentor who said 'draw the failure' made reliability planning concrete. In ML practice (thread 20): We learned that naming a single owner for public circle SEO snippets prevents contradictory descriptions in search results helpfully. The architecture review that asked about multi-region assumptions caught naive defaults.
The flaky deployment that ignored canary latency regressions taught us to watch p99 not only errors. In ML practice (thread 20): The migration that chunked batches avoided locking tables longer than maintenance windows. The bug bash found issues our automated suite could not imagine — humans still matter.
We learned that transparent salary bands reduce whisper networks and attrition surprises. In ML practice (thread 20): Pairing on the scary migration reduced my anxiety more than any document. We learned that transparent pricing for paid circles beats hidden surcharges when hosts compare platforms quarterly helpfully finally always.
We should have deleted unused Lambda functions still billed monthly — small leaks add up. In ML practice (thread 20): Remote made async communication non-optional; emoji tone-misreads were real incidents. We should have invested in backup restore drills before the auditor asked for proof.
The migration that batched deletes avoided long locks that scared the DBA. In ML practice (thread 20): The architecture diagram was wrong but useful until it was dangerously wrong. The smallest improvement to CSV column order matched analyst muscle memory and won hearts.
Join the conversation.
Log in to reply