On 2025-08-14 at 14:22:18, Junio C Hamano wrote: > "brian m. carlson" <sandals@xxxxxxxxxxxxxxxxxxxx> writes: > > I do not know if you want my input (as I wasn't CC'ed), but anyway... > > > ... We can store them in the > > `loose-object-idx`, but since it's not sorted or easily searchable, it's > > going to perform really terribly when we store enough of them. Right > > now, we read the entire file into two hashmaps (one in each direction) > > and we sometimes need to re-read it when other processes add items, so > > it won't take much to make it be slow and take a lot of memory. > > > > For these reasons, I think we need a different datastore for this and > > I'd like to solicit opinions on what that should look like. Here are > > some things that come to mind: > > I do not see why loose-object-idx is not sorted in the first place, > but to account for new objects getting into the object store, it > would not be a viable way forward to maintain a single sorted file. > We obviously do not want to keep rewriting it in its entirety all > the time, It's not sorted because there's no way to do so and efficiently handle both lookups. If we sorted it in SHA-256 order, then we would still have to look up items in SHA-1 order with a linear search, and vice versa. What we do for pack index v3 is a sorted table of abbreviated names, a mapping of that order to pack order, and then full object names in pack order, with a set for each algorithm. The abbreviated names all use the same prefix size, which is just long enough to be unambiguous. This means that we can easily look up an object, find its index into pack order, and then find the full object ID in any algorithm. We could probably write some sort of data file that contains these same mappings except that since we don't have a pack order, we could just use a sorted order in the main algorithm and omit the main algorithm's mapping table. We could then have a single table for the necessary object metadata. > If there aren't any radically novel idea, I would imagine that our > design would default to have a big base file that is optimized for > reading and searching, plus another format that is easier and > quicker to write that would overlay, possibly in a way similar to > packed and loose refs work? Yeah, that could be an option. Or we could have a base file and some incrementals, with a `git gc` when we hit 50 items, just like when we hit 50 packfiles. > As there are some objects for which we need to carry dynamic > information, e.g. "we expect not to have this in our object store > and that is fine", which may be set for objects immediately behind > the shallow-clone boundary, may need to be cleared when the depth of > shallowness changes. Would it make sense to store these auxiliary > pieces of information in separate place(s)? I suspect that the > objects that need these extra bits of information form a small > subset of all objects that we need to have the conversion data, so a > separate table that is indexed into using the order in the main > table may not be a bad way to go. My plan is to just wire this up to `git gc`. We'd know what entries are potentially disposable (such as shallows) and omit the unneeded entries when repacking. -- brian m. carlson (they/them) Toronto, Ontario, CA
Attachment:
signature.asc
Description: PGP signature
