The master proxy was too slow to erase a GRV budget deficit if no GRV requests were coming in. #1999
Conversation
|
I'm adding some additional logic to clamp the budget deficit to this PR, that should be coming shortly. |
…ds of transactions. Decay the deficit if the rate changes and it exceeds the new limit.
|
I've added the new logic. The knob values were chosen somewhat arbitrarily, so let me know if you think they should change. |
# Conflicts: # documentation/sphinx/source/release-notes.rst
| void reset(double elapsed) { | ||
| limit = std::min(0.0,limit) + std::min(rate * elapsed, SERVER_KNOBS->START_TRANSACTION_MAX_TRANSACTIONS_TO_START); | ||
| limit = std::min(0.0, limit) + rate * elapsed; // Adjust the limit based on the full elapsed interval in order to properly erase a deficit | ||
| limit = std::min(limit, rate * SERVER_KNOBS->START_TRANSACTION_BATCH_INTERVAL_MAX); // Don't allow the rate to exceed what would be allowed in the maximum batch interval |
There was a problem hiding this comment.
Is this necessary? It seems like elapsed will not be larger than SERVER_KNOBS->START_TRANSACTION_BATCH_INTERVAL_MAX and thus std::min(0.0, limit) + rate * elapsed will always be a value smaller or equal to rate * SERVER_KNOBS->START_TRANSACTION_BATCH_INTERVAL_MAX.
There was a problem hiding this comment.
Oh never mind, I didn't see the change below.
|
@fdb-build test CTest please |
…of seconds of transactions. Decay the deficit if the rate changes and it exceeds the new limit." This reverts commit 90cb73d.
|
I've reverted the code that clamps the deficit for now because I think there are some flaws both with the current approach and the new one I had implemented. My intent is to try to address those in a future PR. |
3 participants
No description provided.