Re: [RFC PATCH 3/3] list-objects-filter: implement composite filters

[Matthew DeVore <matvore@comcast.net>]

> On May 16, 2019, at 8:25 PM, Junio C Hamano <gitster@pobox.com> wrote:
>> 
>> 	$ git rev-list --filter=tree:2 --filter:blob:limit=32k
> 
> Shouldn't the second one say "--filter=blob:limit=32k" (i.e. the
> first colon should be an equal sign)?

That's right. Fixed locally.

> 
>> Such usage is currently an error, so giving it a meaning is backwards-
>> compatible.
> 
> Two minor comments.  
> 
> If combine means "must satisfy all of these", '+' is probably a poor
> choice (perhaps we want '&' instead).  Also, it seems to me that

I think I agree. & is more intuitive.

> having to worry about url encoding and parsing encoded data
> correctly and securely would be far more work than simply taking
> multiple command line parameters, accumulating them in a string
> list, and then at the end of command line parsing, building a
> combined filter out of all of them at once (a degenerate case may
> end up attempting to build a combined filter that combines a single
> filter), iow just biting the bullet and do the "potentially be
> improved" step from the beginning.

My intention actually is to support the repeated flag pretty soon, but I only want to write the code if there's agreement on my current approach.

My justification for the URL-encoding scheme is:

1. The combined filters will eventually have to travel over the wire.

2. The Git protocol will either have repeated "filter" lines or it will continue to use a single filter line with an encoding scheme.

3. Continuing to use a single filter line seemed the least disruptive considering both this codebase and Git clones like JGit. Other clones will likely fail saying "unknown filter combine:" or something like that until it gets implemented. A paranoid consideration is that clones and proprietary server implementations may currently allow the "filter" line to be silently overridden if it is repeated.

4. Assuming we *do* use a single filter line over the wire, it makes sense to allow the user to specify the raw filter line as well as have the more friendly UI of repeating --filter flags.

5. If we use repeated "filter" lines over the wire, and later start implementing a more complete DSL for specifying filters (see Mercurial's "revsets") the repeated-filter-line feature in the protocol may end up becoming deprecated and we will end up back-pedaling to allow integration of the "&" operator with whatever new operators we need.

(I very much doubt I will be the one implementing such a DSL for filters or resets, but I think it's a possibility)

> So why are we allowing %3A there that does not even have to be
> encoded?  Shouldn't it be an error?

We do have to require the combine operator (& or +) and % be encoded. For other operators, there are three options:

1. Allow anything to be encoded. I chose this because it's how I usually think of URL encoding working. For instance, if I go to https://public-inbox.org/git/?q=cod%65+coverage in Chrome, the browser automatically decodes the %65 to an e in the address bar. Safari does not automatically decode, but the server apparently interprets the %65 as an e. I am not really attached to this choice.

2. Do not allow or require anything else to be encoded.

3. Require encoding of a couple of "reserved" characters that don't appear in filters now, and don't typically appear in UNIX path names. This would allow for expansion later. For instance, "~&%*+|(){}!\" plus the ASCII range [0, 0x20] and single and double quotes - do not allow encoding of anything else.

4. Same requirements as 3, but permit encoding of other arbitrary characters.

I kind of like 3 now that I've thought it out more.

> 
> In any case, I am not quite convinced that we need to complicate the
> parameters with URLencoding, so I'd skip reviewing large part this
> patch that is about "decoding".

It's fine if we drop the encoding scheme. I intentionally tried to limit the amount of work I stacked on top of it until I got agreement. Please let me know if anything I've said changes your perspective.

> 
> Once the combined filter definition is built in-core, the code that
> evaluates the intersection of all conditions seems to be written
> sanely to me.

Great! I actually did simplify it a bit since I sent the first roll-up.

Thanks.