Stelios Grigoriadis wrote:
Howard Chu wrote:
Stelios Grigoriadis wrote:
The output from one run. I haven't had the time to look at it yet.
The key question is, do any of these debug messages correspond to entries that you've found are missing on the replicas?
Entry cn=dlt30604-0,dc=example,dc=com CSN 20071005135356Z#00000e#00#000000 greater than snapshot 20071005135356Z#00000d#00#000000
The entries that are missing from the replicas are the ones with "greater than snapshot". I commented out the if-statement around line 1869 in syncprov.c and couldn't reproduce the problem.
On a side note, if an exception would be made for refresh and persist and send results greater than the snapshot:ed ctxcsn, what are the side effects of this? Is it possible to get "ghost" entries in the replicas? Suppose that an add operation fails to commit but the entry is already replicated, is that possible?
No, not possible. The backends will not return uncommitted entries in search responses.
The only possible side-effect is that the replicas may receive some entries twice if they are picked up during both the refresh phase and during the persist phase. That is, it will consume a little extra network bandwidth.
I'm looking at restructuring the search code to try to avoid this duplication.