Re: Bug#881692: command-not-found: I re-wrote command-not-found

2017-11-14 Thread Shawn Landden 
On Nov 14, 2017 8:15 AM, "Julian Andres Klode"  wrote:

On Tue, Nov 14, 2017 at 03:35:02PM +, John Lenton wrote:
> On 14 November 2017 at 12:34, Julian Andres Klode  wrote:
> > On Tue, Nov 14, 2017 at 01:00:54PM +0100, Zygmunt Krynicki wrote:
> >> I would love if we have a compact representation of mapping from name
> >> to list of bits of information where each bit can be a small structure
> >> with some data. Apart from components for ubuntu archive it could be
> >> used to store facts about snap packages, flatpaks, etc. I would try to
> >> avoid a simplistic command -> package mapping as that will force us to
> >> encode things into strings in an ad-hoc way.
> >
> > That makes sense to me. But then we're back on a db, I guess. I sort
> > like this minimal approach.
>
> I was thinking in the other direction, was going to see how it behaved
> with sqlite as the store. Would that be objectionable?

Using a relational database for a simple key -> structure mapping seems
overkill and a mismatch for the problem, and the SQL does not make it
more readable.

I'd play with lmdb and kyotocabinet, these are two high-performance
key-value file databases and then encode a structure as mentioned
before.

I had some kyotocabinet code, (i maintain that package, which btw is in
mentors) but this way is at least half the size. (Kyotocabinet is 1mb and
it almost doubles the size of the db, even using lower overhead b-tree back
end. These entries are just very small.

For the text file approach, we can even go human, readable, like git:

git just encodes a number in a fixed-length decimal number, we can do
the same, and then just encode (length, key), (length, data) pairs after
each other (or as mentioned, just use the "index" as the field id, and
store field ids in the progrma). Uses a bit more space, but encodes
everything in a format you could read with a text editor, and should
not be terribly less efficient.

The thing is: This needs to be as efficient as possible: it should
be below 100ms (or better 50ms), regardless of whether caches are dropped
or not.

Python code |   Shawn's code

SSD, cache  50ms5ms
SSD, " dropped 256ms   15ms
HDD, cache  50ms5ms
HDD, " dropped 530ms   15ms

I guess Shawn's code could even be improved in performance by
avoiding the subprocess execution, avoiding various ld cache
lookups and library loads.

I am going to have to bring it in process to add the spell check code.


That said, space requirements might matter too.
--
Debian Developer - deb.li/jak | jak-linux.org - free software dev
Ubuntu Core Developer  de, en speaker
-- 
Ubuntu-devel-discuss mailing list
Ubuntu-devel-discuss@lists.ubuntu.com
Modify settings or unsubscribe at: 
https://lists.ubuntu.com/mailman/listinfo/ubuntu-devel-discuss

Re: Bug#881692: command-not-found: I re-wrote command-not-found

2017-11-14 Thread Julian Andres Klode 
On Tue, Nov 14, 2017 at 03:35:02PM +, John Lenton wrote:
> On 14 November 2017 at 12:34, Julian Andres Klode  wrote:
> > On Tue, Nov 14, 2017 at 01:00:54PM +0100, Zygmunt Krynicki wrote:
> >> I would love if we have a compact representation of mapping from name
> >> to list of bits of information where each bit can be a small structure
> >> with some data. Apart from components for ubuntu archive it could be
> >> used to store facts about snap packages, flatpaks, etc. I would try to
> >> avoid a simplistic command -> package mapping as that will force us to
> >> encode things into strings in an ad-hoc way.
> >
> > That makes sense to me. But then we're back on a db, I guess. I sort
> > like this minimal approach.
> 
> I was thinking in the other direction, was going to see how it behaved
> with sqlite as the store. Would that be objectionable?

Using a relational database for a simple key -> structure mapping seems
overkill and a mismatch for the problem, and the SQL does not make it
more readable.

I'd play with lmdb and kyotocabinet, these are two high-performance
key-value file databases and then encode a structure as mentioned
before.

For the text file approach, we can even go human, readable, like git:

git just encodes a number in a fixed-length decimal number, we can do
the same, and then just encode (length, key), (length, data) pairs after
each other (or as mentioned, just use the "index" as the field id, and
store field ids in the progrma). Uses a bit more space, but encodes
everything in a format you could read with a text editor, and should
not be terribly less efficient.

The thing is: This needs to be as efficient as possible: it should
be below 100ms (or better 50ms), regardless of whether caches are dropped
or not.

Python code |   Shawn's code

SSD, cache  50ms5ms
SSD, " dropped 256ms   15ms
HDD, cache  50ms5ms
HDD, " dropped 530ms   15ms

I guess Shawn's code could even be improved in performance by
avoiding the subprocess execution, avoiding various ld cache
lookups and library loads.

That said, space requirements might matter too.
-- 
Debian Developer - deb.li/jak | jak-linux.org - free software dev
Ubuntu Core Developer  de, en speaker

-- 
Ubuntu-devel-discuss mailing list
Ubuntu-devel-discuss@lists.ubuntu.com
Modify settings or unsubscribe at: 
https://lists.ubuntu.com/mailman/listinfo/ubuntu-devel-discuss

Re: Bug#881692: command-not-found: I re-wrote command-not-found

2017-11-14 Thread John Lenton 
On 14 November 2017 at 12:34, Julian Andres Klode  wrote:
> On Tue, Nov 14, 2017 at 01:00:54PM +0100, Zygmunt Krynicki wrote:
>> I would love if we have a compact representation of mapping from name
>> to list of bits of information where each bit can be a small structure
>> with some data. Apart from components for ubuntu archive it could be
>> used to store facts about snap packages, flatpaks, etc. I would try to
>> avoid a simplistic command -> package mapping as that will force us to
>> encode things into strings in an ad-hoc way.
>
> That makes sense to me. But then we're back on a db, I guess. I sort
> like this minimal approach.

I was thinking in the other direction, was going to see how it behaved
with sqlite as the store. Would that be objectionable?

-- 
Ubuntu-devel-discuss mailing list
Ubuntu-devel-discuss@lists.ubuntu.com
Modify settings or unsubscribe at: 
https://lists.ubuntu.com/mailman/listinfo/ubuntu-devel-discuss

Re: Bug#881692: command-not-found: I re-wrote command-not-found

2017-11-14 Thread Julian Andres Klode 
On Tue, Nov 14, 2017 at 01:00:54PM +0100, Zygmunt Krynicki wrote:
> Hey everyone.
> 
> Thank you for your interest in command-not-found.
> 
> On Tue, Nov 14, 2017 at 8:50 AM, Julian Andres Klode  wrote:
> > (forwarding this to ubuntu-devel-discuss and Zygmunt)
> >
> > On Mon, Nov 13, 2017 at 10:33:39PM -0800, Shawn Landden wrote:
> >> Package: command-not-found
> >> Severity: wishlist
> >>
> >> I re-wrote command-not-found to get rid of the python dependancy, and
> >> to reduce the database size, as to reduce memory usage.
> >>
> >> https://github.com/shawnl/command-not-found
> >>
> >> I was preparing to upload it to mentors as command-not-found-ng
> >
> > I also rewrote it years ago, but using the same database format,
> > just in C. It was a lot faster. I don't understand the memory usage
> > bit - it should not matter how large the database is, it's memory
> > mapped, and not read into memory, as such memory usage should be
> > roughly constant.
> >
> > Questions/Comments for your approach:
> >
> > * Did you test your format on a slow HDD with caches dropped? It
> >   must not be slower than the Python one (that one is way too slow
> >   already) - I did, it seems to be faster (0.4 vs 0.68 seconds)
> >   - I believe the database-based C rewrite was even much faster,
> >   though.
> 
> > * update-command-not-found should use apt-get indextargets
> 
> > * You don't store components, hence you cannot tell people to enable
> >   component. That's a very important use case for Ubuntu, where
> >   not all components are enabled by default, but the database is
> >   shipped in the package.
> >
> >   You could just append / to each package name I think,
> >   and strip it away when displaying.
> 
> I would love if we have a compact representation of mapping from name
> to list of bits of information where each bit can be a small structure
> with some data. Apart from components for ubuntu archive it could be
> used to store facts about snap packages, flatpaks, etc. I would try to
> avoid a simplistic command -> package mapping as that will force us to
> encode things into strings in an ad-hoc way.

That makes sense to me. But then we're back on a db, I guess. I sort
like this minimal approach. An approach of course would be to store
a key/value map after the package, something like:

filepackage name
followed by multiple:
key
value

where lengths are 32-bit (16 bit?) integers.

Should not be too hard. Alternatively, this also works

filepackagename
   for each field
 value  for each field

and then you can index stuff

offset(attr_i) = offset(attr_i) + attrs[i]

Lots of options to extend.

> 
> > * You should use getopt_long() to parse command-line options, and
> >   support -h, --help :)
> 
> > * pts_lbsearch belongs into usr/lib/..., not usr/share/...
> >
> > * You don't implement a closest matches function:
> >
> > $ command-not-found thunderbrd
> > No command 'thunderbrd' found, did you mean:
> >  Command 'thunderbird' from package 'thunderbird' (main)
> > thunderbrd: command not found
> > $ ./command-not-found thunderbrd
> > thunderbrd: command not found
> >
> >This one is really important. People do make typos or misremember
> >command names, so the tool needs to be able to deal with that
> 
> +1 on this, the function should be not too hard to implement in C.
> 
> >Should be easy to implement though, although you might hav
> 
> > * You need to Conflict with command-not-found and not Break AFAIUI
> 
> Ideally, to ease the transition, you should do something about the
> python APIs. If yo can keep them (either as pure-python bindings or
> just as a compatible implementation) that would be a plus. If you want
> to drop them then please announce that and see if anything rdepends on
> it.

Oh, hmm.

> 
> > * You do have to Depend on apt-file, as that configures apt to download
> >   the Contents files
> 
> I didn't look at the details but I (hope) this is a build dependency
> and this will be processed somewhere on the archive side.

That's a Debian-only dependency forced upon us by ftpmaster, on Ubuntu 
we can ship the data in the package (or preferably a separate 
command-not-found-data source package).


> 
> > I think it's a worthwhile approach, and I can see it replacing
> > command-not-found if those tiny issues have been fixed. Then you
> > could also avoid the -ng moniker, and just take over the main
> > package (if Zygmunt does not mind), which also avoids a month
> > long NEW process :)
> 
> Yes, though I'd like to participate as we're working on
> command-not-found improvements in snapd and would like to have
> something that fits Debian, Ubuntu as well as (eventually but not
> conflicting at least) Fedora and openSUSE (at least the snapd part).

I'd like that.

-- 
Debian Developer -

Re: Bug#881692: command-not-found: I re-wrote command-not-found

2017-11-14 Thread Zygmunt Krynicki 
Hey everyone.

Thank you for your interest in command-not-found.

On Tue, Nov 14, 2017 at 8:50 AM, Julian Andres Klode  wrote:
> (forwarding this to ubuntu-devel-discuss and Zygmunt)
>
> On Mon, Nov 13, 2017 at 10:33:39PM -0800, Shawn Landden wrote:
>> Package: command-not-found
>> Severity: wishlist
>>
>> I re-wrote command-not-found to get rid of the python dependancy, and
>> to reduce the database size, as to reduce memory usage.
>>
>> https://github.com/shawnl/command-not-found
>>
>> I was preparing to upload it to mentors as command-not-found-ng
>
> I also rewrote it years ago, but using the same database format,
> just in C. It was a lot faster. I don't understand the memory usage
> bit - it should not matter how large the database is, it's memory
> mapped, and not read into memory, as such memory usage should be
> roughly constant.
>
> Questions/Comments for your approach:
>
> * Did you test your format on a slow HDD with caches dropped? It
>   must not be slower than the Python one (that one is way too slow
>   already) - I did, it seems to be faster (0.4 vs 0.68 seconds)
>   - I believe the database-based C rewrite was even much faster,
>   though.

> * update-command-not-found should use apt-get indextargets

> * You don't store components, hence you cannot tell people to enable
>   component. That's a very important use case for Ubuntu, where
>   not all components are enabled by default, but the database is
>   shipped in the package.
>
>   You could just append / to each package name I think,
>   and strip it away when displaying.

I would love if we have a compact representation of mapping from name
to list of bits of information where each bit can be a small structure
with some data. Apart from components for ubuntu archive it could be
used to store facts about snap packages, flatpaks, etc. I would try to
avoid a simplistic command -> package mapping as that will force us to
encode things into strings in an ad-hoc way.

> * You should use getopt_long() to parse command-line options, and
>   support -h, --help :)

> * pts_lbsearch belongs into usr/lib/..., not usr/share/...
>
> * You don't implement a closest matches function:
>
> $ command-not-found thunderbrd
> No command 'thunderbrd' found, did you mean:
>  Command 'thunderbird' from package 'thunderbird' (main)
> thunderbrd: command not found
> $ ./command-not-found thunderbrd
> thunderbrd: command not found
>
>This one is really important. People do make typos or misremember
>command names, so the tool needs to be able to deal with that

+1 on this, the function should be not too hard to implement in C.

>Should be easy to implement though, although you might hav

> * You need to Conflict with command-not-found and not Break AFAIUI

Ideally, to ease the transition, you should do something about the
python APIs. If yo can keep them (either as pure-python bindings or
just as a compatible implementation) that would be a plus. If you want
to drop them then please announce that and see if anything rdepends on
it.

> * You do have to Depend on apt-file, as that configures apt to download
>   the Contents files

I didn't look at the details but I (hope) this is a build dependency
and this will be processed somewhere on the archive side.

> I think it's a worthwhile approach, and I can see it replacing
> command-not-found if those tiny issues have been fixed. Then you
> could also avoid the -ng moniker, and just take over the main
> package (if Zygmunt does not mind), which also avoids a month
> long NEW process :)

Yes, though I'd like to participate as we're working on
command-not-found improvements in snapd and would like to have
something that fits Debian, Ubuntu as well as (eventually but not
conflicting at least) Fedora and openSUSE (at least the snapd part).

Best regards
ZK

-- 
Ubuntu-devel-discuss mailing list
Ubuntu-devel-discuss@lists.ubuntu.com
Modify settings or unsubscribe at: 
https://lists.ubuntu.com/mailman/listinfo/ubuntu-devel-discuss

Re: Bug#881692: command-not-found: I re-wrote command-not-found

2017-11-14 Thread Julian Andres Klode 
(forwarding this to ubuntu-devel-discuss and Zygmunt)

On Mon, Nov 13, 2017 at 10:33:39PM -0800, Shawn Landden wrote:
> Package: command-not-found
> Severity: wishlist
> 
> I re-wrote command-not-found to get rid of the python dependancy, and
> to reduce the database size, as to reduce memory usage.
> 
> https://github.com/shawnl/command-not-found
> 
> I was preparing to upload it to mentors as command-not-found-ng

I also rewrote it years ago, but using the same database format,
just in C. It was a lot faster. I don't understand the memory usage
bit - it should not matter how large the database is, it's memory
mapped, and not read into memory, as such memory usage should be
roughly constant. 

Questions/Comments for your approach:

* Did you test your format on a slow HDD with caches dropped? It
  must not be slower than the Python one (that one is way too slow
  already) - I did, it seems to be faster (0.4 vs 0.68 seconds)
  - I believe the database-based C rewrite was even much faster,
  though.
* update-command-not-found should use apt-get indextargets
* You don't store components, hence you cannot tell people to enable
  component. That's a very important use case for Ubuntu, where
  not all components are enabled by default, but the database is
  shipped in the package.

  You could just append / to each package name I think,
  and strip it away when displaying.
* You should use getopt_long() to parse command-line options, and
  support -h, --help :)
* pts_lbsearch belongs into usr/lib/..., not usr/share/...

* You don't implement a closest matches function:

$ command-not-found thunderbrd
No command 'thunderbrd' found, did you mean:
 Command 'thunderbird' from package 'thunderbird' (main)
thunderbrd: command not found
$ ./command-not-found thunderbrd
thunderbrd: command not found

   This one is really important. People do make typos or misremember
   command names, so the tool needs to be able to deal with that

   Should be easy to implement though, although you might have to
   search multiple times - once for each alternative. All you need is

def similar_words(word):
""" return a set with spelling1 distance alternative spellings

based on http://norvig.com/spell-correct.html"";
alphabet = 'abcdefghijklmnopqrstuvwxyz-_'
s = [(word[:i], word[i:]) for i in range(len(word) + 1)]
deletes= [a + b[1:] for a, b in s if b]
transposes = [a + b[1] + b[0] + b[2:] for a, b in s if len(b)>1]
replaces   = [a + c + b[1:] for a, b in s for c in alphabet if b]
inserts= [a + c + b for a, b in s for c in alphabet]
return set(deletes + transposes + replaces + inserts)

And search for what that returns. And you don't need to search for those
at all if you have a direct match.

* It needs to be translated - also very important.

* You need to Conflict with command-not-found and not Break AFAIUI

* You should not depend on grep, sed, coreutils, they are Essential.

* You do have to Depend on apt-file, as that configures apt to download
  the Contents files

* You should not have identifiers starting with _ in the program, these
  are reserved for the C implementation (like _cleanup_free_).

Yes, and these are basically the same reasons my C prototype is
not in the archive. Also, I did not put a lot of work into it, as
I was waiting for PackageKit to take that over, but that was not
done yet.

I think it's a worthwhile approach, and I can see it replacing
command-not-found if those tiny issues have been fixed. Then you
could also avoid the -ng moniker, and just take over the main
package (if Zygmunt does not mind), which also avoids a month
long NEW process :)

-- 
Debian Developer - deb.li/jak | jak-linux.org - free software dev
Ubuntu Core Developer  de, en speaker

-- 
Ubuntu-devel-discuss mailing list
Ubuntu-devel-discuss@lists.ubuntu.com
Modify settings or unsubscribe at: 
https://lists.ubuntu.com/mailman/listinfo/ubuntu-devel-discuss