Teodor Sigaev <teodor@sigaev.ru> writes:
> Ok, I already realized WAL in GiST and tested with online backup. It works.
> Now I need test it on recovery after crash (power loss etc) and there is a
> problem in GiST with incompleted inserts, I hope, I resolved it, but this
> problem, I think, has the same nature as incompleted splits in btree code. For
> solving, btree xlog code uses stack of incompleted splits and if they leaves
> incompleted, btree_xlog_cleanup function completes them. I make use similar
> idea, but I can't reproduce situation when xlog_cleanup makes some useful work
> except just looks at void stack. So question: how is it possible to reproduce
> desired behaviour?
It's not very easy; you have to arrange for the WAL history to end
between the two action records that you are interested in. What I'd
suggest is- run test backend under GDB, set breakpoint between making log entry for first split step and entry for
second.-let it reach the breakpoint- in another backend, commit some unrelated transaction to force XLOG buffers to
disk-in GDB, "kill -9" to prevent test backend from proceeding
If you want to trace through the recovery as well, the best bet might be
to temporarily add a "sleep(30)" near the beginning of StartupXlog so
that you have time to reattach to the recovery process.
regards, tom lane