If slapd terminates uncleanly, a checkpoint will be lost on the accesslog db. Depending on the syncprov overlay checkpoint settings (usually no checkpointing is enabled on the accesslog db) this can cause the system to refuse engage in replication at startup.
Need to determine the correct method of handling this, to be reviewed again for beta.
syncprov might be able to tell the backend not to commit its TXN on applying the modification, then send the contextCSN update in the result callback. Would remove some of the concerns about performance and I/O impact that has caused us to do contextCSN checkpoints.