[PATCH 02 of 15 V2] bookmark: add methods to binary encode and decode bookmark values

Augie Fackler raf at durin42.com
Mon Nov 13 18:30:39 EST 2017


> On Nov 13, 2017, at 05:16, Boris Feld <boris.feld at octobus.net> wrote:
> 
> On Fri, 2017-11-10 at 17:35 -0500, Augie Fackler wrote:
>> (+indygreg, who also is a formats enthusiast)
>> 
>> On Thu, Nov 02, 2017 at 02:17:59PM +0100, Boris Feld wrote:
>>> # HG changeset patch
>>> # User Boris Feld <boris.feld at octobus.net>
>>> # Date 1508072395 -7200
>>> #      Sun Oct 15 14:59:55 2017 +0200
>>> # Node ID 4d0c6772a81aa1e2b25f32f944563db1f33fd327
>>> # Parent  8c9a9eecdcd61401a1604a08a5272f7dabd4b912
>>> # EXP-Topic b2.bookmarks
>>> # Available At https://bitbucket.org/octobus/mercurial-devel/
>>> #              hg pull https://bitbucket.org/octobus/mercurial-deve
>>> l/ -r 4d0c6772a81a
>>> bookmark: add methods to binary encode and decode bookmark values
>>> 
>>> Coming new bundle2 parts related to bookmark will use a binary
>>> encoding. It
>>> encodes a series of '(bookmark, node)' pairs. Bookmark name has a
>>> high enough
>>> size limit to not be affected by issue5165. (64K length, we are
>>> well covered)
>> 
>> I'm not thrilled here. Could we do some sort of varint encoding,
>> which
>> would be generally useful going forward for a variety of things,
>> rather than just shrugging and setting a limit which we (today) deem
>> absurdly large?
> 
> We are not sure what the value of that would be. Bundle2 makes it very
> easy to move to a newer encoding version if needed.

I mean, not really. We have to add a b2cap, and then we've got to maintain multiple versions forever. It's kind of a pain. A varint encoding wouldn't be hard, and will probably take O(an hour) to code up. I'd entertain that much more eagerly than just dropping a uint64 in here and asserting it should be big enough forever.

> We are already
> doing so for changegroup and obsmarkers.
> 
> The biggest bookmark we have seen was about 800 characters, that should
> leave us some room for the future without blocking us.

That you know about. I seem to remember violent arguments that 255 bytes should be more than enough. I agree a 64k bookmark sounds totally insane, but what if I wanted to put base64'd test result metadata in a bookmark?

> 
>> 
>> I agree that practically speaking, nobody should have a 64k bookmark,
>> but we can do better. We also know that \0 shouldn't appear in a
>> bookmark name, so we could just make the format be
>> null-terminated. I'd much rather we get in the practice of
>> deliberately crafting formats that are maximally flexible and
>> reusable.
> 
> That would be significantly harder to parse since we would need to
> search for the '\0' in the stream. We do not think it brings much value
> either since any new data we sneak into the last field would have to be
> understood by the receiver. If we need to negotiate for such
> capabilities, this is equivalent to rolling out a new encoding.

Then let's explore the varint option. Here's a rough sketch of a reader:

import io
src = io.BytesIO(b'\xff\x01')
cur = ord(src.read(1))
i = cur & 0x7f
while cur & 0x80:
    i = i << 7
    cur = ord(src.read(1))
    i += cur & 0x7f
assert 0b11111110000001 == i

Gives you 7 bits of value per byte transmitted, which means that we'll get most realistic bookmark lengths in two bytes of "here's how much string" (probably one byte for the vast majority of ascii cases!)

I'd much rather we just build up the infrastructure for stuff like this _now_ and start using it, rather than continue to punt and shove massive uint64 fields in and then pray it'll be enough for all eternity. It does mean we can't use struct anymore, but that seems like a comparatively small price to pay for being able to trivially make this future proof.

(As before, if other reviewers think I'm being too hard on this, please speak up.)


More information about the Mercurial-devel mailing list