pathInfoCache: Respect disk cache TTLs #3398 (2.3 backport)

[roberth]

Backport of #3403 without C++20

[edolstra]

Hm, this might be a bit too big of a chance for the stable branch... It's also more of a feature rather than a bug fix.

[roberth]

Caching more would be risky, but this caches less.

narinfo-cache-negative-ttl

The TTL in seconds for negative lookups. If a store path is queried from a substituter but was not found, there will be a negative lookup cached in the local disk cache database for the specified duration.

It does not mention an additional long-lived cache in nix-daemon that can't be cleared.

If the implementation doesn't conform to the docs, isn't that a bug?

[edolstra]

Nix doesn't have any long-lived processes itself though, so it's not really a bug.

[roberth]

Isn't nix-daemon a long-lived process that is part of Nix?

[edolstra]

All the real work is done by the nix-daemon child processes which exist only as long as the connection.

[roberth]

Will Nix 2.4 be released for NixOS 20.09?

[Ericson2314]

@roberth you mean 3.0? I think we could make that.

[roberth]

@Ericson2314 branch-off for 20.09 is in about a week. Is that too soon or is the goal to stabilize it during the release month?

[roberth]

I still have to ask users to patch their system Nix, while this patch has run in production for over half a year. Can it be merged?

[roberth]

Thanks!