On Thu, Mar 31, 2022 at 10:50 AM Junio C Hamano <gitster@xxxxxxxxx> wrote: > > Neeraj Singh <nksingh85@xxxxxxxxx> writes: > > > Just to understand your feedback better, is it a problem to separate > > the state of each separate "thing" under ODB transactions into > > separate file-scope global(s)? In this series I declared the fsync > > state as completely separate from the packfile state. That's why I > > was thinking of it as more of a naming problem, since the remaining > > state aside from the plugged boolean is entirely packfile related. > > Ahh, sorry, my mistake. > > I somehow thought that you would be making the existing "struct > bulk_checkin_state" infrastructure to cover not just the object > store but much more, perhaps because I partly mistook the motivation > to rename the structure (thinking again about it, since "checkin" is > the act of adding new objects to the object database from outside > sources (either from the working tree using "git add" command, or > from other sources using unpack-objects), the original name was > already fine to signal that it was about the object database, and > the need to rename it sounded like we were going to do much more > than the object database behind my head). > > > My argument in favor of having separate file-scoped variables for each > > 'pluggable thing' would be that future implementations can evolve > > separately without authors first having to disentangle a single > > struct. > > That is fine. Would the trigger to "plug" and "unplug" also be > independent? As you said elsewhere, the part to harden refs can > piggyback on the existing ref-transaction infrastructure. I do not > know offhand what things other than loose objects that want "plug" > and "unplug" semantics, but if there are, are we going to have type > specific begin- and end-transaction? > With regards to bulk-checkin.h, I believe for simplicity of interface to the callers, there should be a single pair of APIs for plug or unplug of the entire ODB regardless of what optimizations happen under the covers. For eventual repo-wide transactions, there should be a single API to initiate a transaction and a single one to commit/abort the transaction at the end. We may still also want a flush API so that we can make the repo state consistent prior to executing hooks or doing something else where an outside process needs consistent repo state. Thanks, Neeraj
