Alex, I didn't want to forget this one, so I'm opening an issue.

I'm confused why *clsm_update and *clsm_remove check the cursor's WT_CURSTD_OVERWRITE flag.

The code in each that does:

        if (F_ISSET(cursor, WT_CURSTD_OVERWRITE) ||
            (ret = __clsm_search(cursor)) == 0)

should maybe be:

        if ((ret = __clsm_search(cursor)) == 0)

but that's bad for __clsm_remove, because LSM searches are more expensive.

Is there a reason we can't change the semantics of WT_CURSOR::remove to ignore the possibility the key doesn't exist, and then LSM doesn't have to do a search first.

If that isn't OK, we could at least make that the default and add a new config option to return failure if the key doesn't exist?

This isn't an issue for the Basho/Riak tests, I'm told all of its get/put/del cursors are setting overwrite.