"D1 will create read-only clones of your data, close to where your users are, and constantly keep them up-to-date with changes."

Sounds like there will be no synchronous replication and instead there will be a background process to "constantly keep [read-only clones] up-to-date". This means that a stale read from an older read replica can occur even after a write transaction has successfully committed on the "primary" used for writes.

So, while the consistency is not "thrown away", it's no longer a strong consistency? Anyway, Kyle from Jepsen will figure it out soon, I guess :)

Yeah, so you can always opt-in to strong consistency by transferring execution to the primary (see the "Embedded Compute" section of the blog). Then it's pretty much exactly the same as a DO.

Just clarifying - D1 without read replicas is strongly consistent. If you add read replicas, those can have replication lag and will not be strongly consistent.

Disclaimer: I work at Cloudflare :)

Thanks for the clarification, that is what I would expect.

Does SQLite support some kind of monotonic transaction id that can be used as a cache coherency key? Say a client writes a new record to the database which returns `{"result": "ok", "transaction_id": 123}`, then to ensure that subsequent read requests are coherent they provide a header that checks that the read replica has transaction_id >= 123 and either waits for replication before serving or fails the request. (Perhaps a good use for the embedded worker?)

Since it's a relational DB, and supports transactions, you can have a journal table right?

I know of a very important system at AWS that did this with MySQL :D

Yes you could do it manually, but it would be nice if the solution didn't require carefully managing update queries so the journal addition isn't missed and increasing write amplification to manually update a journal table when that information probably already exists somewhere in the WAL implementation.

Yup sorry about that. I missed the entire "read replica" bit when reading that blog post.

Interesting that D1 is built on top of Durable Objects. Does this mean that it would be practical for a single worker to access multiple D1 databases, so it could use, for example, a separate database for each tenant in a B2B SaaS application? Edit: And could each database be in a different primary region?

Yes, exactly!

That is interesting. I wish CF would give us some more information as I've assumed that there must be a lack of strong consistency which would be a major drawback.

Edit: But that would mean that durable objects can't be replicated asynchronously? That would mean a big latency hit. Then what's the difference to a central DB in one datacenter?

I’m not familiar with Durable Objects. When D1 does replication to read replicas, if it’s not doing synchronous replication, then it’s not strongly consistent, is that correct?

I wish the post had provided some more technical details.

It's more of a "quickstart" than a peek under the hood.

I'd like to see some up front D1 & R2 benchmarks (read/write/iops). I can't judge invocation cost value until I can judge my use case performance. Here's hoping its -gt NVMe raid 10 under the hood of D1 as some big SQLite reads suffer under slow storage.

Are you guys using litestream or a similar approach? E.g storing WAL frames in a durable object.