Github user suyogmapara commented on the issue:
https://github.com/apache/zookeeper/pull/567
Hi @maoling
I tested the patch with following parameters:
1. PreAllocSize = 100 KB, TxnLogSizeLimit = 200 KB, 100 transactions of
little over 10 KBs. Outcome: 5 transaction files with little over 200 KBs each
(rounded up to last transaction).
2. PreAllocSize = 100 KB, TxnLogSizeLimit = 100 KB, 100 transactions of
little over 10 KBs. Outcome: 100 transaction files one per each transaction
3. PreAllocSize = 100 KB, TxnLogSizeLimit = 50 KB, 100 transactions of
little over 10 KBs. Outcome: 100 transaction files one per each transaction.
Second and third outcomes are expected as every time it tries to
pre-allocate more, we would be rolling the log. Of course those won't be
sensible configurations, but if someone uses them, only performance would be
impacted and not the correctness. Please let me know if you disagree.
@anmolnar
snapCount doesn't give direct handle on the transaction log size. Changing
average transaction size from 1 KB to 100 KB would increase the transaction log
size by 2 orders of magnitude and may push they sync time over the edge. More
importantly reducing the snapCount would also increase how frequently we
snapshot and so we would be forced to find a sweet spot to balance the two
conflicting goals. On the other hand, txnLogSizeLimit provides cheap and
independent control on the worst case seek time during the diff sync.
---
