On Mon, Sep 05, 2011 at 02:32:40PM +1000, Greg Banks wrote: > Ok, I'm now convinced my first attempt at annotation quotas sucked too > hard, here's how I want to re-implement it. Let me know what you think.
> - add a 32b mailbox index header entry to track the storage in bytes > used by all annotations for the mailbox itself or for messages in the > mailbox Why not 64b? Admittedly hanging gigabytes of annotations off a single folders is probably evil, but this is the mailbox header - 4 bytes per mailbox to not even have to think about it is super-cheap. We keep a handful of "blanks" around in the header already. > - at header upgrade time, write the special value ~0 into this field, > meaning "unknown" Funky. If you didn't want to get all special-value about it, you could use a flag over in the flags field as well, there are only about 5 of the 32 in use so far. > - in mailbox_commit_quota(), if the field is not "unknown", then > calculate the delta in usage and apply to the quota db. I assume this is done the same way it's done now? By taking a "snapshot" of the value at the start, and calculating a diff to apply at the end. That was actually a filthy workaround (one of many, dammit) for using unsigned datatypes. If we made this a signed 64 bit datatype as well, with a flag off to the side for "unknown", then we could actually store a diff in the mailbox object directly - and update both this field and the quota's value during the commit. > - make reconstruct scan the annotations db to figure out the correct > value for this new field. Yep - that makes sense. Reconstruct already does this for the quota_mailbox_used field, doesn't it. And any other derived "aggregate" values in the header. > This means that the annotation STORE and SETMETADATA paths will be > updating the quota db in the same place that APPEND et al do, > mailbox_commit_quota(). This should work around Bug #3529 and also make > the code neater. Nice. Bron.