Date | Commit message (Collapse) |
|
Decrement regen_down when visiting messages that appear in %D that we
know will later be deleted. This ensures consistent message numbers are
generated no matter which commit number is on top. Allowing deletes to
propagage separately from the messages they delete without causing
problems.
The v2 trees already do this and when the indexes are deleted and
rebuilt they maintain they commit numbers.
Add a v1 version of the v2reindex test to verify that reindexing is
working properly on v1 as well as v2.
Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
|
|
Xapian v1.2.21..v1.2.24 (inclusive) use OFD locks but failed to
set the close-on-exec flag on those locks. So we must continue
to work around those old versions by ensuring Xapian file
descriptors aren't held any longer than necessary when in
long-running git processes.
Reported-by: Konstantin Ryabitsev <konstantin@linuxfoundation.org>
|
|
Xapian will replace files upon committing, so non-parallel
V2Writable users need to have umask preserved this way.
|
|
"LIKE" in SQLite (and other SQL implementations I've seen) is
expensive with nearly 3 million messages in the archives.
This caused some partial Message-ID lookups to take over 600ms
on my workstation (~300ms on a faster Xeon). Cut that to below
under 30ms on average on my workstation by relying exclusively
on Xapian for partial Message-ID lookups as we have in the past.
Unlike in the past when we tried using Xapian to match partial
Message-IDs; we now optimize our indexing of Message-IDs to
break apart "words" in Message-IDs for searching, yielding
(hopefully) "good enough" accuracy for folks who get long URLs
broken across lines when copy+pasting.
We'll also drop the (in retrospect) pointless stripping of
"/[tTf]" suffixes for the partial match, since anybody who
hits that codepath would be hitting an invalid message ID.
Finally, limit wildcard expansion to prevent easy DoS vectors
on short terms.
And blame Pine and alpine for generating Message-IDs with
low-entropy prefixes :P
|
|
I was using this to trace the path of brian's message.
Fixes: 017fed7bc4d33ac4
("searchidx: regenerate and avoid article number gaps on full index")
|
|
Relaxing this lock during a v1 --reindex is important to keep
messages showing up in -watch process in a timely manner.
Looks like I deleted an extra line when doing the following
for v2:
s/xdb->commit_transaction/self->commit_txn_lazy/
Fixes: 35ff6bb106909b1c ("replace Xapian skeleton with SQLite overview DB")
|
|
This increases indexing time by around 10% but roughly
halves memory usage of an -index process.
We will probably make this tunable in the future for people
with bigger/smaller machines.
|
|
We do not want phrase searches to cross between independent
fields (filenames/Message-ID vs bodies)
|
|
Some messages to git@vger went missing from Msgmap from old bugs
and became inaccessible via NNTP. Forcing NNTP article numbers
when the overview DB came about made the problem more visible when
reindexing old (v1) repositories as all removed spam messages
took up AUTOINCREMENT numbers again before they were removed.
Having large gaps in NNTP article numbers is not good since it
throws off NNTP clients. This does NOT prevent NNTP clients from
seeing some messages twice, but is better than having them
miss several messages entirely.
We also avoid depending on --reverse in git-log, as
git requires storing an entire commit list in memory for
--reverse, so it's cheaper to store only deleted blobs in the %D
hash since they do not live long.
|
|
I'm not sure how useful this view is, but it exists for now.
|
|
Otherwise articles show up again...
|
|
Since we only query the SQLite over DB for OVER/XOVER; do not
need to waste space storing fields To/Cc/:bytes/:lines or the
XNUM term. We only use From/Subject/References/Message-ID/:blob
in various places of the PSGI code.
For reindexing, we will take advantage of docid stability
in "xapian-compact --no-renumber" to ensure duplicates do not
show up in search results. Since the PSGI interface is the
only consumer of Xapian at the moment, it has no need to
search based on NNTP article number.
|
|
Since the overview stuff is a synchronization point anyways,
move it into the main V2Writable process and allow us to
drop a bunch of code. This is another step towards making
Xapian optional for v2.
In other words, the fan-out point is moved and the Xapian
partitions no longer need to synchronize against each other:
Before:
/-------->\
/---------->\
v2writable -->+----parts----> over
\---------->/
\-------->/
After:
/---------->
/----------->
v2writable --> over-->+----parts--->
\----------->
\---------->
Since the overview/threading logic needs to run on the same core
that feeds git-fast-import, it's slower for small repos but is
not noticeable in large imports where I/O wait in the partitions
dominates.
|
|
Favor simpler internal APIs this time around, this cuts
a fair amount of code out and takes another step towards
removing Xapian as a dependency for v2 repos.
|
|
Dscho found this useful for finding matching git commits based
on AuthorDate in git. Add it to the overview DB format, too;
so in the future we can support v2 repos without Xapian.
https://public-inbox.org/git/nycvar.QRO.7.76.6.1804041821420.55@ZVAVAG-6OXH6DA.rhebcr.pbec.zvpebfbsg.pbz
https://public-inbox.org/git/alpine.DEB.2.20.1702041206130.3496@virtualbox/
|
|
This allows us to emulate the display of thread-aware MUAs when
multiple messages share the same Message-ID. This also is a
place where "public-inbox-index --reindex" is useful to fix
existing messages and no schema version bump is necessary.
|
|
This ought to provide better performance and scalability
which is less dependent on inbox size. Xapian does not
seem optimized for some queries used by the WWW homepage,
Atom feeds, XOVER and NEWNEWS NNTP commands.
This can actually make Xapian optional for NNTP usage,
and allow more functionality to work without Xapian
installed.
Indexing performance was extremely bad at first, but
DBI::Profile helped me optimize away problematic queries.
|
|
We can store :bytes and :lines in doc_data since we never
sort or search by them. We don't have much use for the Date:
stamp at the moment, either.
|
|
We will vivify multiple ghosts if a message has multiple
Message-IDs.
|
|
We'll be making sure V2Writable uses this.
|
|
We do not need this subroutine for read-only use in Search.pm
|
|
This should help us detect bugs sooner in case we have
space waste problems.
|
|
This still requires a msgmap.sqlite3 file to exist, but
it allows us to tweak Xapian indexing rules and reindex
the Xapian database online while -watch is running.
|
|
We want to rely on Date: to sort messages within individual
threads since it keeps messages from git-send-email(1) sorted.
However, since developers occasionally have the clock set
wrong on their machines, sort overall messages by the newest
date in a Received: header so the landing page isn't forever
polluted by messages from the future.
This also gives us determinism for commit times in most cases,
as we'll used the Received: timestamp there, as well.
|
|
While parallel processes improves import speed for initial
imports; they are probably not necessary for daily mail imports
via WatchMaildir and certainly not for public-inbox-init. Save
some memory for daily use and even helps improve readability of
some subroutines by showing which methods they call remotely.
|
|
This reduces code duplication needed for locking and
and hopefully makes things easier to understand.
|
|
We need to hide removals from anybody hitting the search engine.
|
|
Followup-to: ebb59815035b42c2
("searchidx: do not modify Xapian DB while iterating")
|
|
We'll let the config of all.git dictate every other subrepo to
ease maintenance and configuration. The "include" directive has
been supported since git 1.7.10, so it's safe to depend on as v2
requires git 2.6.0+ anyways for "get-mark" in fast-import.
|
|
We can't rely on header order for Message-ID after all
since we fall back to existing MIDs if they exist and
are unseen. This lets us use SearchMsg->mid to get the
MID we associated with the NNTP article number to ensure
all NNTP article lookups roundtrip correctly.
|
|
Since we support duplicate MIDs in v2, we can safely truncate
long MID terms in the database and let other normal duplicate
resolution sort it out. It seems only spammers use excessively
long MIDs, and there'll always be abuse/misuse vectors for causing
mis-threaded messages, so it's not worth worrying about
excessively long MIDs.
|
|
Since we support duplicate MIDs in v2, the NNTP article number
becomes the true unique identifier and we want a way to do fast
lookups on it.
While we're at it, stop putting XPATH in the term partitions
since we only need it in the skeleton DB.
|
|
Aside from the Message-Id ('Q'), these terms do not appear in
content and thus have no business contributing to the Xapian
document length.
Thanks-to Olly Betts for the tip on xapian-discuss
<20180228004400.GU12724@survex.com>
|
|
When indexing diffs, we can avoid indexing the diff parts under
XNQ and instead combine the parts in the read-only search
interface. This results in better indexing performance and
10-15% smaller Xapian indices.
|
|
It's possible to have a message handle multiple terms;
so use this feature to ensure messages with multiple MIDs
can be found by either one.
|
|
'Q' is merely a convention in the Xapian world, and is close
enough to unique for practical purposes, so stop using XMID
and gain a little more term length as a result.
|
|
It's shorter and more convenient, here.
|
|
This is a bit expensive in a multi-process situation because
we need to make our indices and packs visible to the read-only
pieces.
|
|
|
|
Iterating through a list of documents while modifying them does
not seem to be supported in Xapian and it can trigger
DatabaseCorruptError exceptions. This only worked with past
datasets out of dumb luck. With the work-in-progress "v2"
public-inbox layout, this problem might become more visible
as the "thread skeleton" is partitioned out to a separate,
smaller Xapian database.
I've reproduced the problem on both Debian 8.x and 9.x with
Xapian 1.2.19 (chert backend) and 1.4.3 (glass backend)
respectively.
|
|
Interchangably using "all", "skel", "threader", etc. were
confusing. Standardize on the "skeleton" term to describe
this class since it's also used for retrieval of basic headers.
|
|
We will need timestamp, YYYYMMDD, article number, and line count
for querying thread information (including XOVER for NNTP).
|
|
This used to lookup the message in git, but no longer, so
remove a needless indirection layer and call add_message
directly.
|
|
It works around some bugs in older Email::MIME which we'll
find useful.
|
|
This should give us an idea of how much a problem deduplication
will be.
|
|
The parallelization requires splitting Msgmap, text+term
indexing, and thread-linking out into separate processes.
git-fast-import is fast, so we don't bother parallelizing it.
Msgmap (SQLite) and thread-linking (Xapian) must be serialized
because they rely on monotonically increasing numbers (NNTP
article number and internal thread_id, respectively).
We handle msgmap in the main process which drives fast-import.
When the article number is retrieved/generated, we write the
entire message to per-partition subprocesses via pipes for
expensive text+term indexing.
When these per-partition subprocesses are done with the
expensive text+term indexing, they write SearchMsg (small data)
to a shared pipe (inherited from the main V2Writable process)
back to the threader, which runs its own subprocess.
The number of text+term Xapian partitions is chosen at import
and can be made equal to the number of cores in a machine.
V2Writable --> Import -> git-fast-import
\-> SearchIdxThread -> Msgmap (synchronous)
\-> SearchIdxPart[n] -> SearchIdx[*]
\-> SearchIdxThread -> SearchIdx ("threader", a subprocess)
[* ] each subprocess writes to threader
|
|
This is too slow, currently. Working with only 2017 LKML
archives:
git-only: ~1 minute
git + SQLite: ~12 minutes
git+Xapian+SQlite: ~45 minutes
So yes, it looks like we'll need to parallelize Xapian indexing,
at least.
|
|
In general, they are, but there's no way for or general purpose
mail server to enforce that. This is a step in allowing us
to handle more corner cases which existing lists throw at us.
|
|
I decided not to copy the notmuch implementation regarding
serialization of integers to Xapian metadata.
|
|
This will allow easier-compatibility with v2 code which will
introduce content_id as the unique identifier.
The old "XMID" becomes "XM" as a free text searchable term.
"Q" becomes "XMID" as a boolean prefix.
There's no user-visible changes in this, but there needs to
be a schema version bump later on...
(more changes planned which can affect v1)
|