Logging Performance


  1. Don’t slow the user down

    1. logging operations like commit and pull need to run as though logging wasn’t there (or very close).

  2. Avoid saturation of the server by minimizing times locks are held.

  3. ?


From 7/19/02 meeting of Andrew, Wayne and Rick Andrew summarizes (Rick’s edits):

a) We need Wayne’s null sync fix on bitkeeper.com  — can we upgrade bitkeeper.com to the bk-2.1.x TOT ? This pending rick/lm fix the MYSQL issues…​

Note: I tried a more aggressive version of Wayne’s fix, takepatch/resolve cannot handle it. (so only tested one is the original null sync fix)

b) revisit making push/pull propagate log maker When pulling in a change that has already been logged, and it becomes the tip, there is no need to log. This could happen often if people use BK to grab latest code, but never do changes. It can reduce log connection traffic for null sync.

c) push part2 lock retry - do we want it on server side or client side ? Currently client. client side fix has transitional issues: we do not want the server and client to both do the retry.

d) keysync protocol change.. Reduce the amount of data that gets sent back from logging tree. First, quantify problem to make sure when this is worth fixing.

e) x.lmarker a way to make the log 2 probe data more meaningful in the case of the trunk already being probed after a merge. Then can send the probe along the branch instead. Advantage: no protocol or server change, only client. Disadvantage: Not quantified how much this will help.