Hiroshi Inoue <Inoue@tpf.co.jp> writes:
> My concern is the robustness of rel cache.
> It seems pretty dangerous to discard system relation
> descriptors used for cache mechanism especially in
> case of error recovery.
> It also seems pretty dangerous to recontruct relation
> descriptors especially in case of error recovery.
Why? We are able to construct all the non-nailed relcache entries
from scratch during backend startup. That seems a sufficient
proof that we can reconstruct any or all of them on demand.
Until the changes I made today, there was a flaw in that logic,
namely that the specific order that relcache entries are built in
during startup might be somehow magic, ie, building them in another
order might cause a recursive syscache call. But now, that doesn't
matter.
regards, tom lane