Dawid Ciężarkiewicz

PRs would be appreciated. Otherwise it might wait for me having more time. :)

There could be many lock files, each with a timestamp and explicit expiration. An agent that is holding the lock would be supposed to write out a new lock file...

> Using remote cloud storage as KV store is bound to be extremely slow I don't understand this one. Is it because KV store doesn't have a rename (changing the...

> The store is locked during this enumeration, so no new names or chunks can be created while creating and recording the list. I don't think you understand the original...

The only weakness of the current scheme that I'm aware of is that no data can be ever mutated in place, so things have to change names This is not...

> But, in the scenario you describe, you can in fact end up with more than two generations in flight if both (or several) synchronized hosts start GC cycles independently...

I doubt there's going to be one strategy for all use-cases, but my bet would be that 2, max 3 would be all that is ever needed, and they would...

I'm open for having such thing integrated directly, but why does it have to use a naming scheme? The creation date etc. could be stored in the name file itself,...

``` $ cat foo/0000000000000000-c39fdb79bc3faa16/name/foo.yml --- digest: d202d7951df2c4b711ca44b4bcc9d7b363fa4252127e058c1a910ec05b6cd038 index_level: 0 ``` Please add any metadata you want that is missing (date mostly?), while backfilling date during deserializatoin with filesystem creation/modification date...

> That would mean that a prune command must read each and every name file, instead of just doing a ls. Is that wanted? Doesn't seem terrible, especially that the...