Engineering, priced to the unit. Independent · Source-linked · Reproducible

Buildopsy

Case studies connect what teams have publicly disclosed to what the architecture costs per request, task, or event. Facts, inference, and estimates stay distinct; models expose their assumptions so you can rerun them.

Editorial standards and accountability · About Buildopsy

50+ Teardowns
Postmortems & RFCs
30 Tools
Free browser calculators
$0
Articles, courses, and tools

The Buildopsy method

Start with cost per successful request →

Featured Analysis

View All RFCs
Cutaway schematic of the Inngest database path: a client row with an expired deadline, a PgBouncer waiting queue, a gate admitting one database connection at a time, and a rack of connection slots marked closed rather than returned, with 44, 48 and 55 minute incident windows below
Featured Postmortem Durable Execution

Inngest Postmortem: Connections Died Faster Than the Pooler Could Replace Them

A lookup deadline expired while its query waited inside PgBouncer. The driver cancelled and closed the client connection; the pooler later ran the abandoned query and closed that backend connection instead of returning it. Both poolers refill one connection at a time.

InngestPgBouncerPostgreSQLPostmortem
Read the postmortem

Recent Engineering Writing

Browse All Articles
Break-even chart comparing the cost of repeating a prompt prefix with no cache, a 30-minute cache write, and a 12-hour cache write
AI InfrastructureOctober 20268 min read

Prompt Caching Economics: How Many Reuses a Cache Write Has to Earn

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.

Prompt CachingLLM CostsPricing
Read the teardown
Schematic of the europe-west2-b power chain: a grid feed, a UPS battery bank marked drained, two generators with both PLC controllers failed, the dark zone rack, and the surviving zone racks taking failover demand while their control-plane calls time out, with 6h 10m, 5h 29m and 41m durations below.
Cloud InfrastructureOctober 202611 min read

Google Cloud's London Outage: One Zone Lost Power, the Region Lost Capacity

Both controllers that start the backup generators failed, one zone went dark, and the region failed in two ways: failover demand saturated the survivors, and control-plane calls still had to reach metadata in the dead zone.

Google CloudAvailability ZonesPowerPostmortem
Read the postmortem
Starlink core-network failure: satellite constellation not identified as the failed layer, failed ground core services, and the exposed 6M+ user footprint; tens of thousands reported offline
Satellite NetworkOctober 202619 min read

Starlink's 2.5-Hour Global Outage: The Core Network Failed, Not the Satellites

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.

StarlinkControl PlaneOutageCore Network
Read the postmortem
Illustrative x-ray sketch of a Roadster-style coupe with the reported cold-gas components called out — external nitrogen fill, COPV tanks, regulator and valves, rear nozzles, and a refill loop — with reported durations beside each configuration: about five seconds of boost and a 2–6 hour refill for the 200-liter Thrust, a 6–7 hour refill for the 600-liter Rocket. A dashed panel holds a modeled 80-kilonewton, five-second gas-flow scenario whose inputs come from separate reports.
Vehicle EngineeringOctober 2026

Reported 0.81 Seconds to 60, Hours to Refill: The Roadster's Cold-Gas Math

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.

TeslaCold GasClaims AuditRook
Read the audit
Reported sequence: suspicious lookalike book listing, author purchase and report, marketplace removal
AI Book Slop AuditOctober 2026

Amazon's AI Book Slop Problem: The Copycat Beside the Real Book

A reporter bought a suspicious lookalike of his own book. Amazon removed it after his report; the seller and generation method remain unverified.

AmazonPublishingProvenance
Read the audit
Editorial card: the LiteLLM v1.107.0 tag was clean, but the build came from a lookalike fork; a Rook supply-chain audit.
AI Supply Chain IncidentOctober 2026Rook

Ten Attacker Commits Above the Last Clean Commit: The LiteLLM 1.107.0 Supply-Chain Audit

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.

LiteLLMSupply ChainCI/CDRook
Read the audit
Editorial illustration separating the OpenCode Console, where provider keys may have been exposed, from the local app; OpenCode reported 1.4% of Console users impacted.
AI Security IncidentOctober 2026

OpenCode Console Breach: The Missing Link in the Distillation Claim

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.

OpenCodeAPI KeysDisclosure Audit
Read the audit
Sprites' October 6 Grok Meta read: 3,349 ads, 1,526 running, a calculated 45.6% active share, and conflicting 35- and 36-creative read counts
Advertising SystemsOctober 2026

The Grok Meta Ad Audit: 3,349 Ads, Two Creative Denominators

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.

Grok AdsMeta Ad LibrarySprites
Read the audit
Two panels: a GPT-6.1 Sol case where the monitor kept running but human alerts were suppressed, and a GPT-6 Astra case where the monitor was removed while conversation logging stayed on
AI Security IncidentOctober 2026

OpenAI's Agents Weakened Their Own Safety Monitor — and One Denied It

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.

OpenAIMonitoringAgent Safety
Read the postmortem
Timeline of OpenRouter's Feb 17 and Feb 19, 2026 outages: cache drops database connections, recovers cold, database saturates, users see 401 User not found
Cloud InfrastructureOctober 2026

OpenRouter's February 2026 Outages: How a Cold Cache Overwhelmed the Database

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.

OpenRouterCacheThundering HerdPostmortem
Read the postmortem
Diagram of two Workers during Cloudflare R2's March 2025 credential rotation: the default Worker receives the new credential while production keeps the old one until 22:45
Cloud InfrastructureOctober 2026

Cloudflare R2's 67-Minute Outage: Wrong Worker, Old Credentials

A secret written without --env landed on the default Worker. When Cloudflare deleted the old R2 credential, 100% of writes failed for 67 minutes.

CloudflareR2Credential RotationPostmortem
Read the postmortem
Azure OpenAI incident flow: cache-clearing deployment, about 20× normal regional-backend traffic, exhausted connections, and database-dependent health-check restarts
AI InfrastructureOctober 2026

Azure OpenAI Sweden Central: ~20× Normal Backend Traffic, 5h55 of Impact

Microsoft's PIR traces cache misses, connection exhaustion and a database-dependent restart loop in Sweden Central.

Azure OpenAICachingPostmortem
Read the postmortem
PHP2Shell timeline: three warning intervals before release, from CVE identifiers to Calif’s video to Chainguard’s early images, with the fix and leak lanes
Security Incident AuditOctober 2026

PHP2Shell: The Fix Was Ready. The Shutdowns Happened Anyway.

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.

PHPDisclosureRook
Read the audit
Night illustration of two us-west1 data centers linked by a fiber route: the primary link under maintenance with warning lights and a service truck, an overloaded alternate path carrying a queue of packets with a few dropped, and a status panel reading primary under maintenance and alternate congested
Cloud InfrastructureSeptember 2026

GCP us-west1 Outage: Fiber Maintenance and Failed Rerouting

Google's final report says planned fiber work reduced inter-datacenter capacity and rerouting did not properly redistribute traffic to alternate capacity.

GCPNetwork CapacityPostmortem
Read the postmortem
Uber Eats search pipeline teardown: serial stages with false dependencies redrawn as overlapped parallel lanes
Search InfrastructureOctober 2026

Uber Eats Search: The Above-the-Fold Metric Behind a 50% Latency Cut

They stopped measuring API time and started measuring hunger. Every lever priced: 120 retrieval, 100 hydration, 130 ads, 200 infra.

UberSearchLatency
Read the teardown
GOMEMLIMIT teardown: the container-blind collector's two failure modes and Zomato's dynamic limit loop
Cloud InfrastructureOctober 2026

One Soft Limit, $30K a Month: Zomato's GOMEMLIMIT Teardown

GC CPU fell from up to 60% to under 2% across 250+ Go services. The soft-limit mechanism, priced from Zomato's own engineering report.

GoGOMEMLIMITTeardown
Read the teardown
Egress pricing teardown: AWS charges $0.09 per gigabyte out while R2 charges nothing, over capacity-priced transit
Cloud InfrastructureOctober 2026

The $0.09 Toll: Why Egress Is the Cloud's Real Lock-In

One terabyte out of S3 costs $90 in transfer. Out of R2 it costs $0. The volume-priced tollbooth on capacity-priced transit — with worked math.

EgressR2Teardown
Read the teardown
Controlled measurements in Germany and Scotland; one selected single-satellite window rules out satellite switching as the cause of its observed change
Satellite NetworkOctober 2026

Starlink's 15-Second Rhythm: Evidence for a Global Scheduler

Measurements report 15-second allocation and link changes at tested US and European sites. A single-satellite window rules out satellite switching there; the wider mechanism remains unresolved.

StarlinkSchedulerTeardown
Read the teardown
Starlink latency diagram summarizing SpaceX-reported peak-hour median and p99 changes, listed driver categories, and the absence of per-fix contributions
Satellite NetworkOctober 2026

Starlink Latency: What SpaceX Reports From 48.5ms to 33ms

SpaceX reports a US peak-hour median change from 48.5ms to 33ms and lists latency drivers; the whitepaper gives no measured per-fix split.

StarlinkLatencyTeardown
Read the teardown
DPDP compliance data plane: consent receipts, retention ledger, erasure orchestration, and breach clock
Privacy EngineeringOctober 2026

DPDP Compliance Is a Data-Plane Problem: Consent, Erasure, and Breach Math

Consent receipts, erasure across replicas, and breach clocks your on-call owns. India's law, engineered for every jurisdiction.

DPDPGDPRErasure
Read the teardown
Starlink core-network failure: satellite constellation not identified as the failed layer, failed ground core services, and the exposed 6M+ user footprint; tens of thousands reported offline
Satellite NetworkOctober 2026

Starlink's 2.5-Hour Global Outage: The Core Network Failed, Not the Satellites

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.

StarlinkControl PlanePostmortem
Read the postmortem
Identity attack chain behind the Rockstar breach and the court timeline to the November 2026 retrial
Security & IdentityOctober 2026

GTA 6 Leak Postmortem: The Identity Chain Behind Rockstar's Breach

Valid credentials, push-fatigue MFA, and SSO lateral movement put 90 GTA 6 clips on a forum. Retrial November 9, ten days before launch.

Lapsus$MFASSO
Read the postmortem
Datadog observability meters combining into a monthly bill with the Coinbase $65M invoice as scale
ObservabilityOctober 2026

Coinbase's $65M Datadog Bill: Per-Host, Per-GB and Per-Event Meters

Per-host, per-GB, and per-event meters compounded into a $65M invoice. How it happened, why Coinbase stayed, and the guardrails.

DatadogObservabilityCosts
Read the postmortem
CircleCI workflow database WAL saturation stopping dispatch and creating a runner backlog
CI/CD InfrastructureOctober 2026

CircleCI Outage Postmortem: Full WAL Saturation Blocked 3,500 Job Starts

Anti-wraparound autovacuum saturated the orchestration database's WAL volume. The 50-minute outage became hours of queue recovery.

PostgreSQLWALPostmortem
Read the postmortem
Comparison of Deezer-reported synthetic track upload share and listener stream share
AI Music AuditOctober 2026

Deezer's AI Music Flood: 90,000 Tracks a Day, Only 1–3% of Streams

A denominator-by-denominator audit of synthetic music uploads, listener share, Spotify spam disclosures and stream fraud.

AI MusicDeezerStreaming Fraud
Read the audit
Reference architecture showing an agent workload behind a kernel isolation boundary and host-controlled default-deny egress gateway, with write-event monitoring
Agent InfrastructureSeptember 2026

AI Sandbox Escape System Design: Proxy Bypass, Egress Gates and Containment Cost

How proxy exceptions fail and what identity-bound egress controls cost under an explicit workload model.

SandboxProxyEgress
Read article
Discord message path through channel partitions, request coalescing, and the storage layer
Messaging InfrastructureSeptember 2026

How Discord Stores Trillions of Messages: Cassandra, Hot Partitions and ScyllaDB

Channel partitioning, quorum read fan-in, Rust coalescing and the engine swap from 177 nodes to 72.

DiscordCassandraScyllaDB
Read article
Kafka rebalance diagram comparing classic eager revocation, cooperative assignment, and conditional static membership
Event InfrastructureSeptember 2026

Kafka Consumer Rebalance Explained: Storms, Cascades and Fixes

Classic eager versus cooperative assignment, conditional static membership, and a modeled example where one original and two retries means three attempts per affected call.

KafkaRebalancingRetries
Read article
Notion request path through API authorization, PostgreSQL shards, change-data capture, the data lake, and search
Forensic ArchitectureSeptember 2026

Notion Architecture: How 200 Billion Blocks Survive Postgres, Kafka and AI Search

How Notion scales: 480 Postgres shards, CDC lake, vector tiering and the bill behind a blank page.

NotionPostgresKafka
Read article
Inference cost path from retrieval through reranking, prompt prefill, and answer generation
AI InfrastructureSeptember 2026

Perplexity Postmortem: The Answer Engine That Pays for Every Question Twice

A forensic teardown of RAG retrieval, KV cache flaws, prefill tax, 780M monthly queries and the cost difference at 1M events per second.

RAGInference1M RPS
Read article
Diagram of repeated agent install attempts producing separate billed runs
Agent RuntimeSeptember 2026

Replit Agent Postmortem: The Day the Demo Became Production

A deep teardown of prompt-to-production software, sandbox execution, agent retries, deployment risk and the cost difference at 1M events per second.

AI AgentsSandbox1M RPS
Read article
Vercel AI Gateway runtime model comparing active CPU work with time waiting on providers
Cloud InfrastructureSeptember 2026

Vercel Fluid Compute: The Serverless Bill Was Waiting for a Response

A forensic look at concurrency, cold starts, compute-resolver, AI Gateway waiting time and the cost difference at 1M RPS.

ServerlessAI Gateway1M RPS
Read article
Two collaborators make conflicting edits while one document owner orders the resulting change
Realtime SystemsSeptember 2026

Figma Multiplayer: Server-Authoritative Edits and a Modeled 65k RPS Bill

A practical teardown of WebSocket collaboration, Rust isolation, conflict resolution and the hidden price of an always-live canvas.

WebSocketsRustRealtime
Read article
Notion data flow from block edits through sharded Postgres and Kafka CDC to the lake, search and vector tier
Data InfrastructureAugust 2026

Notion's 200 Billion Blocks and the Warehouse Bill Behind a Blank Page

A forensic teardown of Notion's block database, 480 logical shards, CDC pipeline, vector search and modeled cloud bill.

PostgresKafkaS3
Read article
Distributed SystemsJuly 2026

Rethinking Control Planes: Cell-Based Architecture

A control-plane redesign focused on regional isolation, failure containment, Raft and zero-blast-radius operations.

GoRaftAWS
Read article

New field guides