3 ms·
Is Scalable OLTP in the Cloud a Solved Problem? (CIDR 2023)
- bob1029 3y agoI don't think it is for all applications, but for many it might be. We are looking at using Az Hyperscale SQL, but the single writer thing might cause some friction. Trying to figure out if their txn log cap is sufficient for our business at ~1000 customers: > Transaction log throughput cap is set to 100 MB/s for any Hyperscale compute size. The ability to achieve this rate depends on multiple factors, including but not limited to workload type, client configuration and performance, and having sufficient compute capacity on the primary compute replica to produce log at this rate. Having some constraints like this might do good for our engineering moving forward. Today, we are pretty sloppy with data storage representations because every client gets their own dedicated SQLite database and bandwidth/storage are mostly a non-issue at scale. If we knew we had a global limit on DB writes, we'd get way more deliberate with direct schema representations (as opposed to plopping JSON blobs into various "data" columns). Each user interaction today usually costs a few megabytes of IO with SQLite. In theory, we could reduce this to a few hundred bytes per on average.