Quoth Ramakrishnan Muthukrishnan on Aug 05 at 11:35 pm:
> Ramakrishnan Muthukrishnan <ram at rkrishnan.org> writes:
> > Hi,
> > Pardon me for using non-standard terms here because of lack of knowledge
> > of the language used in the email world.
> > I have this thread in which some messages are tagged unread. When the
> > thread is opened, I see that those unread messages are not shown in full
> > in the threaded view. Instead they appear `folded' with only the summary
> > appearing in the thread, as if they are read.
> > Here is a screenshot: http://i.imgur.com/Eh8SKe6.png
> > I am attaching the messages in the thread shown in the screenshot as a
> > tar.bz2 file.
> I just forgot to mention that I see this only when I do a
> notmuch-search with `*' and not when I search with a tag.
'*' turns out to be the key. When showing a thread,
notmuch-show-build-buffer constructs a new query from the thread ID
and the original search query of the form
thread:X and (original query)
If the original query is '*', this is
thread:X and (*)
But * isn't part of the Xapian query syntax. Notmuch specially
handles queries that exactly match "*" before passing the query to
Xapian. When * is embedded in a larger query, this special casing
doesn't apply. In fact, Xapian parses this query as
(Tmail AND (and:(pos=1) FILTER GX))
The "and" in the query turned into a plain search term (the "AND" is a
proper boolean operator, but is unrelated to the provided query).
This is a symptom of a general problem where we assume queries are
textually composable, when they are not. We have the same problem at
least in notmuch-search-filter and in notmuch tag query optimization.
In this particular case, the * causes Xapian's query grammar to fail
to parse the query, which Xapian handles by re-parsing the entire
query will all query features disabled (which includes disabling
support for boolean operators). Unfortunately, just handling * better
isn't really a solution because it's just one of many things that
violates query composability.
Some solutions I can see are:
1) Switch to a composable query syntax (which would include *). This,
obviously, requires a custom query parser, but is the most
localized change I can think of and keeps queries as strings.
2) Never construct queries by pasting strings together. This would
require changes to both the libnotmuch and CLI interfaces and
queries could no longer be strings, but in the words of Alan
Perlis, the string is a stark data structure. (In the case of
show, I would actually love it if we could specify separate search
and match queries because that would eliminate --entire-thread as
well as the fallback in notmuch-show-build-buffer when search
returns nothing, but I digress.)
3) Keep queries as plain strings, but switch to some hybrid syntax
that lets us combine Xapian queries with composable operators
parsed by notmuch. When we need to combine queries, do it using
the composable operators. This actually may not be a bad way to
transition to a full custom query parser; I think it would be
relatively easy to take over parsing Xapian's boolean syntax, but
leave the "prob" parsing to Xapian.
I lean towards 3 because it seems like the least disruptive and offers
a smooth transition to 1.