Optimize struct type validation. The regexp's show in profiles.

[ghost]

While doing something else, I noticed that sync appeared to have gotten slower. I bisected it and discovered that 5e901b0 regressed syncing ~300MB of sf_scrime data from ldb to ldb by ~ 100%.

After some short digging, I'm fairly sure it's because of the addition of type validation to the validating_batch_sink.

Thinking about it some, I think it's an easy fix because of the type cache. Basically we're validating over and over again exactly the same in-memory type graph (the type cache ensures that identical types are interned). we just need to find a way to only do it once per type.

[ghost]

cc @aboodman to set priority

[ghost]

this essentially affects writing of data to a remote server (which logically includes syncing to a local ldb store)

[aboodman]

Let's just drive it by perf of attic. For example I just asked @arv to estimate difficulty of #2869 because it badly affects attic ui perf. If we find that this bug also affects perf, then fix. Otherwise no.

[aboodman]

IOW, not a prio right now.

[ghost]

Blocked by #3123

[arv]

Is this still valid?

[ghost]

I looked recently and I kinda looked like it was. The problem is that the regex is expensive.

[arv]

Which regexp is this? The one that validates the name?

[ghost]

Yup. Struct and field name.