Of Mice & Elephants
summary
A production job writes thousands-deep nested paths and triggers constant failures that, combined with an object-store firmware bug, causes the cluster to be offline for nearly five days while teams copy millions of tiny files out one-by-one. Investigation finds 88% files <5MB, builds a SonarQube plugin to catch bad queries, adds a spark-sampling utility to compute optimal repartitioning, re-architects with object-backed tiers and experiments with Ozone, runs 281 chaos tests, and reports ~96% provisioning and ~60% batch runtime gains.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:29Data Lake
- 7:55Small Files
- 14:05SRE retrospective
- 14:28Data ingestion
- 16:05SRE principles
- 18:00Govermation
- 20:01Leap Engineering
- 24:30Collaboration
- 25:28Learning Development
related talks
01h 06mBest practices when accessing Big Data or any other data!
40mProductionizing Big Data - Stories from the Trenches
25mData Demons: Unleashing Chaos or Driving Performance?
24mTaming Data Skew in Production ML Pipelines
44mAtomic Habits in writing high performing distributed code in the cloud
20mFull Lifecycle Data Analysis on a Large-scale and Leadership Supercomputer...
39mMaturing Your Data Architecture in a Week: How Bluesky Survived
40mTop 10 data engineering mistakes #bbuzz
