Conversation

Franck Pachot

@FranckPachot

Aug 2

I used to explain AWS Aurora as "write as if it is data guard / read as if it is RAC". Really nice to read

@KnownAsMarkus

analysis on on Aurora "The log is the database" architecture. I'll add a few immediate subjective opinions on it👇

Quote Tweet

Markus Michalewicz

@KnownAsMarkus

Aug 2

Based on #AlloyDB's and #AmazonAurora's approach, some of us at #Oracle were asked about the usefulness of applying redo in the storage tier. These are some of our thoughts: blogs.oracle.com/maa/post/about /cc @dbinmemory, @ashishray

Franck Pachot

@FranckPachot

Aug 2

Big difference between AWS Aurora read replicas and Oracle Active Data Guard queries: when you add the cache invalidation arrows onto the Aurora architecture diagram, it quickly looks like a monolithic database spaghetti plate over multiple instances

Franck Pachot

@FranckPachot

Aug 2

>> the same idea has been picked up by two primarily cloud vendors simultaneously Not really simultaneously. GCP is a follower here. And

@neondatabase

is an open-source newcomer. Same idea of separating compute and storage but different times.

Franck Pachot

@FranckPachot

Aug 2

>> Both solutions send the WAL directly to the read replicas, which amplifies writes Not really. WAL is sent to storage. 6 mirrors across 3 AZs. This is the write amplification, but for HA with q quorum of 4. Then read replicas on top, plus cache invalidation.

Franck Pachot

@FranckPachot

Aug 2

Note that this (Aurora & AlloyDB) apply to PostgreSQL with full-row copy on any update and full page logging, on cloud providers where inter-AZ is not free. Oracle has optimized in-place updates, and free inter-AD network cost. Main reason for Aurora was the write cost

Peter Geoghegan

@petervgeoghegan

Replying to

@FranckPachot

The WAL records used for updates in Postgres don't usually include the whole row

4:55 AM · Aug 3, 2022Twitter Web App

Replying to

OK I need to look at it again 🙏 Do you mean 1. When not logging the full page and 2. With HOT updates then WAL record holds only the new value? Or in more general cases?

Replying to

See commit a3115f0d9e, which added the update WAL optimization back in 2014. It's unrelated to FPWs. The optimization applies to HOT and non-HOT updates, though in the case of non-HOT it only works when the update won't move the new version to another page.