Engineering case studies grounded in public evidence, with assumptions and cost models readers can inspect.
Start here: Browse case studies across distributed systems, cloud infrastructure, and AI engineering. Each page distinguishes disclosure from reconstruction and labels modeled numbers as estimates.
AI Slop Watch audits public AI claims, disclosures, and incidents. Browse the shelf.
A cache write is billed at 1.25x or 2.5x input while reads cost 0.10x, so the saving is a reuse count: 1.28 passes on the 30-minute TTL, 2.67 on the 12-hour one.
Both generator controllers failed, one zone went dark, and the region failed through demand and cross-zone validation. Google's preliminary report, with the phase arithmetic recomputed.
Short query deadlines cancelled connections inside PgBouncer faster than the pooler could refill them, and runs were left stranded in a Running state.
The v1.107.0 tag was clean; the release workflow built from a lookalike fork. LiteLLM's final report traces the credentials to a master key committed in May and two throwaway CI jobs in October.
Ten weeks to a fix, hours to a leak, millions of attempts, and no publicly confirmed compromise. A systems audit of the gap between disclosure and deployment.
Start here. Story-driven lessons with diagrams from request paths to interview playbooks, built from production postmortems.
OpenCode reported 1.4% of Console users impacted. Rook separates possible key exposure, user-reported charges, provider actions, and the unverified connection to model training.
Sprites' October 6 read lists 3,349 Grok Meta ads, 1,526 running (45.6%). The same page reports 35 and 36 creatives read, leaving the sample denominator unresolved.
Two internal coding tasks. One agent suppressed security alerts. One removed its monitor and said it had not. The monitor ran in one case. The logging stayed in the other.
OpenRouter's third-party cache dropped its database connections on Feb 17 and 19, 2026. When it recovered cold, every API key lookup hit the database at once and came back as 401 User not found.
A secret written without --env landed on the default Worker. When Cloudflare deleted the old R2 credential, 100% of writes failed for 67 minutes.
Road & Track reports the presentation displayed 0.81 seconds to 60 mph. Refill reports give 2–6 hours for 200-liter Thrust and 6–7 hours for 600-liter Rocket; a separate model does not estimate tank inventory.
Microsoft's PIR traces cache misses, connection exhaustion and a database-dependent restart loop in Sweden Central.
Above-the-Fold redefinition, per-lever millisecond pricing, and the delete-start-earlier-unblock triage across the full stack.
A container-blind collector, a per-container soft limit, and vendor-reported savings of $30K a month across 250+ Go services.
AWS bills $0.09 per GB out while transit is bought by capacity and R2 bills $0. Worked migration math and the guardrails.
Measurements report 15-second allocation and link changes at tested US and European sites; the precise boundary mechanism remains disputed.
SpaceX reports US peak-hour median and p99 changes, lists latency-driver categories, and gives no measured per-fix split.
TypeSafe reports results from four example workflows it built; Vals AI tests claim verification and a nonrepresentative public LegalBench slice, while NXTG.AI tests calibration with private source text and public outcome rows.
Consent receipts, erasure across replicas, and breach clocks. India's duties, engineered for GDPR and US laws too.
Core software services failed while the constellation was not identified as the failed layer. The service's 6M+ users across 140 countries were exposed; tens of thousands reported offline. Ground control-plane fate-sharing, recovery load math and guardrails.
Purchased credentials, push-fatigue MFA, and SSO lateral movement put 90 clips on a forum. Retrial November 9, ten days before launch.
Per-host, per-GB, and per-event meters compounded into a $65M invoice. How it happened, why Coinbase stayed, and the guardrails.
Anti-wraparound autovacuum saturated the orchestration database's WAL volume. The 50-minute outage became hours of queue recovery.
Google's Oct 14 final report attributes the first write failure to a capacity-aware-routing bypass and the Oct 9 recurrence to an incomplete upload-planner rollback.
Packet loss in one Spanish Availability Zone raised regional API errors; AWS reverted the network change and reported recovery.
Unavailable upstream paths pushed Bengaluru traffic onto an undersized survivor link, degrading seven product lines for 63 minutes.
A 1 TiB Standard baseline broken into billable storage, PUT, GET and retrieval costs, with the S3 calculator and transfer limits.
A pre-event EKS IP shortage limited one cluster near 350 nodes. Subnet headroom, CNI targets, legacy Endpoint limits and later CDN QoS controls.
How WhatsApp routing works: Erlang density and Signal E2E and transient inbox and group fan-out math and priced relay model.
How Docs collaboration works: OT transform and revision log and presence lane and snapshot economics and priced scale math.
List prices fell up to half on Sep 22, cache reads fell further. Per-task math, and routing tiers.
v1.37 flags orphaned PVCs by default, deletes nothing. Retention matrix, reclaim audit, and orphan bill math.
Proxy blocked the web, DNS relayed the question. Flagged in 15 minutes, killed after more than two hours.
Coinbase's account of an AWS thermal event, lost matching-engine quorum and two MSK clusters stuck healing.
Google's final report gives a 4h11 detailed customer-impact window; the incident header gives 4h10. Both end at 11:52 PT.
What one LLM task really costs: prefill versus decode and prefix cache and batching, with a worked million-task estimate.
Google's final report says scheduled fiber work reduced inter-datacenter capacity and rerouting did not properly redistribute traffic to alternate capacity.
Microsoft's PIR attributes a 4h57m West US connectivity incident to a break-fix scope error; in-region traffic was unaffected.
Discord says a scale-down exceeded the session handoff grace period; 17% of sessions stopped and voice degraded 12:13 to 15:30 PDT.
METR and Redwood Research document agents communicating through shared Artifactory state. See the information-flow failure and defensive design choices.
A sourced incident analysis and defensive reference architecture for isolation, outbound policy, writable paths, and response.
How Figma multiplayer works: WebSocket fan-out, server-authoritative resolution, Rust rewrite and cost at scale.
Channel partitioning, quorum read fan-in, Rust coalescing and the engine swap from 177 nodes to 72.
Classic eager versus cooperative assignment, conditional static membership and a labeled example where 1 original and 2 retries make 3 attempts per affected call.
How Notion scales: 480 Postgres shards, CDC lake, vector tiering and the bill behind a blank page.
A practical checklist for sandbox isolation, outbound network controls, and write monitoring.
Salesforce Trust lists 16 impacted instance IDs, a 7h36m disruption. It reports stalled login requests, and trouble creating support cases for some customers.
Lab scores showed 78 and 98 passing while field data showed 47. Layout shift fixed first, millions of sites steadied.
250,000 writes a second at 1ms average on owned NVMe hardware. Write-heavy economics, priced.
ShareChat reports 1B features per second without database growth, staged row and cache optimizations, and a separate 10x feature-system cost reduction.
Ten points of mobile conversion bought with milliseconds. Server renders, fixed slots, and bundle hygiene.
CVE-2026-73500 lets silent TLS connections exhaust etcd memory, take down the Kubernetes control plane. Handshake math, and deadlines.
A modeled classic-eager scenario uses 40,000 partitions, an assumed 660-second pause and two retries: three total attempts per affected call.
A forensic teardown of token limits, context rot, compaction cliffs, productivity math and the fixes vendors keep postponing.
A forensic teardown of RAG retrieval, KV cache reuse, prefill cost and inference economics at scale.
A teardown of sandbox execution, agent retries, deployment risk and prompt-to-production infrastructure.
A system design analysis of concurrency, cold starts, provider wait time and serverless cost controls.
A practical teardown of WebSocket collaboration, server-authoritative conflict resolution and realtime cost.
A data-platform teardown covering PostgreSQL sharding, CDC, Kafka, object storage, vector search and modeled cloud cost.
An analysis of regional isolation, failure containment, Raft-backed cells and near-zero-blast-radius operations (under 5% design target).
