In-cluster latency

Sorry couldn’t find this number in the blogs, what’s the latency of distributed (cross-tablet) transactions, 99%, min, mix, avg, within a medium size cluster under moderate load?

hi @bro
This is a complex scenario that depends on multiple factors: concurrency on single rows, % of data in ram, how many tablets and leader servers the query interacts, cpu performance, wal, inter-region/zone,network congestion/latency etc.

As a general point, we can explain how the read/write/transaction path works:

  1. Single key read path https://docs.yugabyte.com/latest/architecture/core-functions/read-path/
  2. Single key write path https://docs.yugabyte.com/latest/architecture/core-functions/write-path/
  3. How isolations levels work/conflict https://docs.yugabyte.com/latest/architecture/transactions/isolation-levels/
  4. Single row transaction https://docs.yugabyte.com/latest/architecture/transactions/single-row-transactions/
  5. Distributed transaction overview https://docs.yugabyte.com/latest/architecture/transactions/distributed-txns/
  6. Distributed transaction read/write path https://docs.yugabyte.com/latest/architecture/transactions/transactional-io-path/

The best way to go about it, is to:

  1. look how your app works in a high level
  2. know the internals so you have an idea to group data better and lower as much the number of network hops
  3. test & repeat depending on the numbers

We can help on all these steps so you can go faster and don’t have to learn all of it.