Philip Martin <philip.mar...@wandisco.com> writes: > I've been thinking about this overnight and I believe I need to do more > investigation on the WANdisco side.
I now realise I was mistaken about this problem. The fact that we don't fsync during a transaction doesn't affect Apache restarts. When Apache acknowledges an http request all the file descriptors to the repository associated with processing that request have been closed and so all data written has moved from user-space to kernel-space. The fact that it is not on disk doesn't matter, it will be visible to other processes. The absence of fsync could make a difference is if the machine, rather than just Apache, were restarted, and I suppose a client/proxy could attempt one commit that spanned a machine restart, but I'm not going to worry about it. The problem that was seen during WANdisco testing occurred when Apache was restarted after a commit had happened in the repository but before Apache could acnowledge the http request. When this happens the transaction associated with the commit may still exist, or may have been partially or fully removed. That's simply normal operation. -- Philip
