Episode 0 is the written overview, not a recording: what we are building, who it is for, and the one picture that explains the stack. Read it before the recordings.
Read19 August 2026
6 of 8 recorded episodes are out. New here? Start with episode 0
A code change alters the schema before replication begins. The Supabase database records what happened, Pipelines moves it, the warehouse models it, and your GTM tools act on the result.
The series is for the person on a Supabase team who gets asked for numbers nobody can produce. Episode 1 covers the chain, episodes 2 to 4 build it, 5 to 7 use it, and 8 is about what breaks it.
Episode 0 is the written overview, not a recording: what we are building, who it is for, and the one picture that explains the stack. Read it before the recordings.
Read19 August 2026
Four layers, and each owns exactly one job. Read them left to right and something uncomfortable falls out: every guarantee downstream is conditional on the layer to its left. Which is why the stack really starts one step before Postgres.
Watch and read6 August 2026
From an empty BigQuery dataset to a table sitting at Live. The publication is SQL you write yourself, the destination takes four fields, and there are two things nobody mentions until you hit them: who gets the bill, and the region field you can never change.
Watch and read14 August 2026
Publishing every table is the wrong answer, and it takes about a month to find out why. Primary keys, replica identity, and the TOAST problem that drops column values from your updates without an error. Plus the tables we leave behind on purpose, and what it would cost us not to.
Watch and read2 September 2026
Replicated tables do not arrive as tables. Each one is a view over a versioned table, and pointing a dashboard at the wrong object works right up until someone truncates the source. Then the real work: turning faithful copies into models that answer something.
Watch and read28 September 2026
First usage starts the clock. One check passes at 10:04 on its first run. Another fails twice before passing at 09:41, 29 minutes after it started. The warehouse model separates activation from first value so you can find accounts that need help.
Watch and read9 October 2026
The 10:04 row has four data sources connected, a last check at 16:20, and six warnings resolved with none ignored. The 09:12 row has one data source, checks that stopped at 09:41, and three warnings ignored. Customer success reads those three facts separately.
Watch and read27 October 2026
The same models as customer success with the comparison reversed. Which accounts have outgrown the tier they are paying for, measured by repository volume and by how many distinct people trigger reviews. Almost the same query, entirely different conversation.
Four ordinary migrations that break the chain without failing CI or turning a pipeline red. A default that does not replicate, history that never backfills, a type change nobody supports, and a truncate that moves the table out from under your dashboard.
Season one, eight episodes
One email per episode. No digest, no drip sequence, no sales cadence. Unsubscribe in one click.
Your address goes on this list and nowhere else.