_1751731246795-BygAaJJK.png){kind=logo}
Saas for Businesses
Playbooks and case studies covering saas for businesses.
How to Migrate a Multi-Tenant SaaS Schema Without Breaking Tenants
Every migration guide assumes all consumers of the database deploy at the same time. In a multi-tenant SaaS, they don't. Here's the playbook we run when tenants are on different release tracks and a single ALTER TABLE would break half your customers.
Backfill Embeddings for 1M Rows Without Killing Postgres
Backfilling embeddings into a live Postgres table with a million rows isn't an ETL job — it's a queue problem. Here's a lease-based pattern that survives crashes, respects rate limits, and never holds a lock longer than one UPDATE.
Kafka vs. SQS: Pick the Right Queue for Your AI Pipeline
Choosing between Kafka and SQS for an AI pipeline usually comes down to one question most benchmarks skip: do you need the message log to be a replayable asset for retraining and audit? Here's the honest tradeoff.
How to Cut Over a Live Database Schema Without Downtime
A step-by-step playbook for renaming columns, dropping tables, and restructuring core models on a live production database — without a maintenance window. Written for CTOs whose deployment pipeline isn't clean enough to ship app and DB changes atomically.
Temporal Tables vs. Audit Logs: Pick the Right History Model
Audit logs and temporal tables solve different problems, and picking the wrong one leaves you writing ad-hoc reconstruction queries for years. Here's how to choose when compliance, point-in-time reporting, or both land on your desk.
AI Observability Cheatsheet: What to Log, Alert, and Ignore
Standard APM misses the signals that actually break LLM features in production. This cheatsheet covers exactly what to log, what to alert on, and what to stop paging engineers about at 2am.
What Is a Webhook? And Why Your Integration Keeps Breaking
A webhook is a push notification between servers — the provider calls you when something happens, instead of you polling them. Most webhook failures in production are not missed deliveries; they are duplicate deliveries processed twice.
Fine-Tune or RAG? Pick the Right AI Memory Strategy
Most fine-tune vs RAG debates treat this as a capability contest. It isn't. The real decision hinges on how fast your knowledge changes, how much labeled data you can produce, and which failure modes you can tolerate.
AI Model Latency Budgets: A Cheatsheet for Product Teams
Latency tolerance isn't a property of your model — it's a property of where the result appears in the user's workflow. A reference for setting defensible p99 targets for AI features in production SaaS.
Feature Store on a Budget: Serve ML Features from Postgres
You don't need Feast, Tecton, or a Redis tier to stop training-serving skew. A properly designed append-only feature table in the Postgres you already run will fix it — here's the exact schema, queries, and gotchas.
LLM Prompt Failure Modes: A Diagnostic Cheatsheet
Most production LLM failures aren't model failures — they're prompt-contract violations with predictable symptoms. A diagnostic reference for engineers debugging summarization, extraction, and classification features that broke after launch.
Celery vs. Kafka for AI Pipelines: Pick the Right One
Celery's broker-ack model makes it structurally incapable of replaying inference events, no matter how much you tune it. Here's how to decide whether your AI pipeline needs a queue or a log — and when Kafka is worth the operational cost.