Catalogue
- Cards
- 165
- Records
- 486
- Workloads
- 19
- Corpus version
- 74847648e25a
76 of 486 records match; the first 20 are shown.
| Kind | Record | Organisation | Year | Setup |
|---|---|---|---|---|
| Failure | Traffic peak overloaded one datacenter's load balancers while a service-mesh sidecar failed to scale, and a client retry bug amplified authentication traffic | GitHub | 2026 | Multi-region |
| Decision | Partition catalog clustering to per-shop tasks and filter standalone stores deterministically via Liquid themes before calling LLMs | Shopify | 2026 | |
| Decision | Replace naive linkage-tree clustering with FAISS HNSW plus sparse UPGMA average linkage to form coherent chunks up to 200 products | Shopify | 2026 | |
| Decision | Use a two-stage propose-and-critique LLM flow with dynamic per-chunk Structured Output schemas and normalized product IDs | Shopify | 2026 | |
| Decision | Own the webhook pipeline after customers could not see what happened to a webhook, leaving a hosted queue | Marble | 2026 | |
| Decision | Workers atomically claim a delivery before sending so overlapping workers or retried queue messages do not send duplicates | Marble | 2026 | |
| Failure | Connection pools that recycled every 15 minutes synchronized into a thundering herd after a managed Postgres upgrade | Clerk | 2025 | |
| Failure | One customer's endpoint outage made deliveries time out, filled the shared worker pool and delayed Directory Sync webhooks for everyone | WorkOS | 2025 | |
| Decision | Pair GPT-4 confirmation and MetaReflection hybrid prompts with expanded offline evaluation to curb LLM false positives | GitHub | 2025 | |
| Decision | Share LLM capacity across push and full-history scans with dynamic rate-limit ranges and filepath exclusions | GitHub | 2025 | |
| Decision | Validate prompt and filter changes via mirror testing on public-preview repositories before GA | GitHub | 2025 | |
| Failure | Users who follow hundreds of thousands of accounts made hot shards that slowed the timelines sharing them | Bluesky | 2025 | Sharded |
| Decision | Drop timeline writes with a probability set by how many accounts a user follows, accepting imperfect timelines | Bluesky | 2025 | Sharded |
| Decision | Cache high-follow accounts in a Redis sorted set loaded into memory every 30 seconds instead of reading follow counts per write | Bluesky | 2025 | Sharded |
| Failure | A surge of duplicate webhooks and the retry setup filled one queue's provider quota, so new events were not delivered for hours | Onfido | 2024 | |
| Decision | Use few-shot GPT-4-chat (gpt-4-32k) instead of fine-tuning davinci-002 or using the completion API | Canva | 2023 | |
| Decision | Redact internal data pre-LLM, use soft 600-character prompt targets with retries, and track edits via Jira webhooks | Canva | 2023 | |
| Decision | Run nearest-neighbor vector search in memory inside stateless Kubernetes services rather than a stateful database | Spotify | 2023 | |
| Decision | Build Voyager as a new HNSW package to succeed Annoy instead of making breaking changes in hnswlib | Spotify | 2023 | |
| Decision | Replace the Wide & Deep (XGBoost + DNN) fraud ensemble with a ResNeXt-inspired multi-branch pure DNN | Stripe | 2023 |
56 more records. Sign in to see all 76.