one-time-iterators

[Austin Clements]

On Thu, May 26, 2011 at 6:22 PM, Patrick Totzke
<patricktotzke at googlemail.com> wrote:
> Excerpts from Austin Clements's message of Thu May 26 22:43:02 +0100 2011:
>> > > Though, Patrick, that solution doesn't address your problem.  On the
>> > > other hand, it's not clear to me what concurrent access semantics
>> > > you're actually expecting.  I suspect you don't want the remaining
>> > > iteration to reflect the changes, since your changes could equally
>> > > well have affected earlier iteration results.
>> > That's right.
>> > > But if you want a
>> > > consistent view of your query results, something's going to have to
>> > > materialize that iterator, and it might as well be you (or Xapian
>> > > would need more sophisticated concurrency control than it has).  But
>> > > this shouldn't be expensive because all you need to materialize are
>> > > the document ids; you shouldn't need to eagerly fetch the per-thread
>> > > information.
>> > I thought so, but it seems that Query.search_threads() already
>> > caches more than the id of each item. Which is as expected
>> > because it is designed to return thread objects, not their ids.
>> > As you can see above, this _is_ too expensive for me.
>>
>> I'd forgotten that constructing threads on the C side was eager about
>> the thread tags, author list and subject (which, without Istvan's
>> proposed patch, even requires opening and parsing the message file).
>> This is probably what's killing you.
>>
>> Out of curiosity, what is your situation that you won't wind up paying
>> the cost of this iteration one way or the other and that the latency
>> of doing these tag changes matters?
>
> I'm trying to implement a terminal interface for notmuch in python
> that resembles sup.
> For the search results view, i read an initial portion from a Threads iterator
> to fill my teminal window with threadline-widgets. Obviously, for a
> large number of results I don't want to go through all of them.
> The problem arises if you toggle a tag on the selected threadline and afterwards
> continue to scroll down.

Ah, that makes sense.

>> > > Have you tried simply calling list() on your thread
>> > > iterator to see how expensive it is?  My bet is that it's quite cheap,
>> > > both memory-wise and CPU-wise.
>> > Funny thing:
>> >  q=Database().create_query('*')
>> >  time tlist = list(q.search_threads())
>> > raises a NotmuchError(STATUS.NOT_INITIALIZED) exception. For some reason
>> > the list constructor must read mere than once from the iterator.
>> > So this is not an option, but even if it worked, it would show
>> > the same behaviour as my above test..
>>
>> Interesting.  Looks like the Threads class implements __len__ and that
>> its implementation exhausts the iterator.  Which isn't a great idea in
>> itself, but it turns out that Python's implementation of list() calls
>> __len__ if it's available (presumably to pre-size the list) before
>> iterating over the object, so it exhausts the iterator before even
>> using it.
>>
>> That said, if list(q.search_threads()) did work, it wouldn't give you
>> better performance than your experiment above.
>>
>> > would it be very hard to implement a Query.search_thread_ids() ?
>> > This name is a bit off because it had to be done on a lower level.
>>
>> Lazily fetching the thread metadata on the C side would probably
>> address your problem automatically.  But what are you doing that
>> doesn't require any information about the threads you're manipulating?
> Agreed. Unfortunately, there seems to be no way to get a list of thread
> ids or a reliable iterator thereof by using the current python bindings.
> It would be enough for me to have the ids because then I could
> search for the few threads I actually need individually on demand.

There's no way to do that from the C API either, so don't feel left
out.  ]:--8)  It seems to me that the right solution to your problem
is to make thread information lazy (effectively, everything gathered
in lib/thread.cc:_thread_add_message).  Then you could probably
materialize that iterator cheaply.  In fact, it's probably worth
trying a hack where you put dummy information in the thread object
from _thread_add_message and see how long it takes just to walk the
iterator (unfortunately I don't think profiling will help much here
because much of your time is probably spent waiting for I/O).

I don't think there would be any downside to doing this for eager
consumers like the CLI.