Eric Wong [Thu, 20 Aug 2020 20:24:51 +0000 (20:24 +0000)]
searchview: convert nested and Atom display to over.sqlite3
git blob retrieval dominates on these, "&x=t" (nested) is
roughly the same due to increased overhead for ->get_percent
storage balancing out the mass-loading from SQLite.
Atom "&x=A" is sped up slightly and uses less memory in the
long-lived response.
Eric Wong [Thu, 20 Aug 2020 20:24:50 +0000 (20:24 +0000)]
searchview: speed up search summary by ~10%
Instead of loading one article at-a-time from over.sqlite3, we
can use SQL to mass-load IN (?,?, ...) all results with a single
SQLite query. Despite SQLite being in-process and having no
network latency, the reduction in SQL query executions from
loading multiple rows at once speeds things up significantly.
We'll keep the over->get_art optimizations from the previous
commit, since it still speeds up long-lived responses, slightly.
Eric Wong [Thu, 20 Aug 2020 20:24:49 +0000 (20:24 +0000)]
searchview: use over.sqlite3 instead of Xapian docdata
This is a step towards improving kernel page cache hit rates by
relying on over.sqlite3 for document data instead of Xapian.
Some micro-optimization to over->get_art was required to
maintain performance.
While this is unlikely to be a problem in current practice,
keeping Xapian DBs open for long responses can interfere with
free space recovery after -compact.
In the future, it will interfere with inbox search grouping
and lead to unexpected results.
Eric Wong [Thu, 20 Aug 2020 20:24:41 +0000 (20:24 +0000)]
search: v2: ensure shards are numerically sorted
This seems required to correctly get the NNTP article number
from Xapian docid on combined Xapian DBs. The default
(ASCII-betical) sorting was only acceptable for -imapd users
until somebody hit 11 (or more) shards, which is a rare case.
Eric Wong [Wed, 19 Aug 2020 08:02:33 +0000 (08:02 +0000)]
smsg: handle wide characters in raw mail headers
There may be messages in the wild with wide characters in
headers which aren't non-RFC2047 encoded. Assume UTF-8 so
those fields can round trip through over.sqlite3.
This doesn't affect docdata.glass in Xapian, but it does
affect how over.sqlite3 stores the same deflated info.
Eric Wong [Thu, 13 Aug 2020 08:04:04 +0000 (08:04 +0000)]
grok-pull.post_update_hook: favor --sequential-shard for HDD
--sequential-shard offers better performance on HDD than -j0
since the on-disk active set can be kept small (with -j $HIGH_NUM).
--batch-size can also be helpful for systems with much RAM.
Eric Wong [Thu, 13 Aug 2020 08:04:37 +0000 (08:04 +0000)]
index|compact|xcpdb: support --all switch
For -index, this is a convenient way to quickly index all
inboxes after a grok-pull. Might as well support it for
rarely used commands like -compact and -xcpdb, too.
Eric Wong [Wed, 12 Aug 2020 09:17:18 +0000 (09:17 +0000)]
xcpdb: wire up new index options and --help
--sequential-shard also disables the copy parallelism (--jobs),
so it can be useful for systems unable to handle parallel random
I/O but still want many shards.
There was a missing "use strict", too, which is fixed.
Eric Wong [Wed, 12 Aug 2020 09:17:16 +0000 (09:17 +0000)]
xapcmd: reduce CPU idling when shards exceeds job count
In case there's unbalanced shards AND we're limiting parallelism
while using many shards, spawn the next task in the queue ASAP
once a task is done, instead of waiting for all tasks to finish
before spawning the next batch.
Unbalanced shards probably isn't a big issue for most users;
however many smaller shards with few jobs can be useful for HDD
users to reduce the effect of random writes.
Eric Wong [Mon, 10 Aug 2020 02:12:04 +0000 (02:12 +0000)]
convert: check ARGV more correctly
Instead of silently ignoring excessive args, don't let a user
specify an extra directory. Furthermore, we'll support the odd
case where BOFH wants to name an $INBOX_DIR to be `0' :P
Eric Wong [Mon, 10 Aug 2020 02:12:00 +0000 (02:12 +0000)]
index: cleanup internal variables
Move away from hard-to-read alllowercase naming and favor
snake_case or separated-by-dashes.
We'll keep `--indexlevel' as-is for now, since it's been around
for several releases; but we'll support `--index-level' in the
CLI and update our documentation in a few months.
We'll also clarify that publicInbox.indexMaxSize is only
intended for -index, and not -watch or -mda.
Eric Wong [Mon, 10 Aug 2020 02:11:57 +0000 (02:11 +0000)]
msgmap: tmp_clone: simplify + meaningful filename
Trying to use the newer ->sqlite_backup_to_dbh method doesn't
seem worth it, as we'll have to support DBD::SQLite <= 1.60
another decade or more.
Dumping 'msgmap-XXXXXXX' into $INBOX_DIR can appear a bit
confusing to users, so give it a "mm_tmp-$PID-XXXXXXXX" name
to emphasize it's a temporary file tied to a given PID.
We also don't want to penalize read-only daemons with
loading File::Temp, so do it lazily.
-index now invokes ->DESTROY like xcpdb does, which is necessary
to cleanup $INBOX_DIR/msgmap-XXXXXXX files. We'll also exit
with the expected values for various signals by adding 128
as described in <https://www.tldp.org/LDP/abs/html/exitcodes.html>
-xcpdb now terminates worker processes and xapian-compact(1)
invocations when prematurely killed, too.
Eric Wong [Mon, 10 Aug 2020 02:11:54 +0000 (02:11 +0000)]
doc: index: more notes about latest changes
With LKML on an HDD, a giant --batch-size of 500m ends up being
pretty useful. I was able to index LKML in ~16 hours on a
system that had other activity on it. The big downside was it
was eating up over 5g of RAM :x.
We'll also fix up a duplicated indexBatchSize section, fix
formatting around global vs per-inbox indexSequentialShard,
and ensure section 5 manpages are linked correctly.
Eric Wong [Mon, 10 Aug 2020 02:11:53 +0000 (02:11 +0000)]
index: --sequential-shard works incrementally
We should never reindex all data in Xapian unless --reindex is
specified on the command-line. This means users who put
publicInbox.indexSequentialShard in their config file won't have
to put up with a full reindex at every invocation, only when
they specify --reindex.
We'll also cleanup the progress output to not emit non-sensical
ranges where the starting number is higher than the end.
Eric Wong [Sat, 8 Aug 2020 11:24:05 +0000 (11:24 +0000)]
favor `getconf _NPROCESSORS_ONLN` over GNU nproc
getconf(1) itself is POSIX, while `_NPROCESSORS_ONLN' is not.
However, FreeBSD (tested 11.4 and 12.1) and glibc (tested CentOS
7.x and Debian 10.x) both support `getconf _NPROCESSORS_ONLN'.
GNU coreutils (and thus `nproc' or `gnproc') are not installed
by default on the *BSDs, so we'll try the option most likely
to exist on both glibc and *BSDs out-of-the-box.
Eric Wong [Sat, 8 Aug 2020 04:59:49 +0000 (04:59 +0000)]
dir_idle: require Perl 5.22+ for kqueue
IO::KQueue requires us to use fileno(DIRHANDLE) for setting up
kqueue watches. This use of fileno() is only supported since
Perl 5.22, so BSD users on older Perl will have to fall back to
old polling.
This affects users of -watch, currently; but will affect other
read-only Xapian users soon.
Eric Wong [Fri, 7 Aug 2020 10:52:18 +0000 (10:52 +0000)]
index: add built-in --help / -?
Eventually, commonly-used commands run by the user will all
support --help / -? for user-friendliness. The changes from
up-front `use' to lazy `require' speed up `--help' by 3x or so.
Eric Wong [Fri, 7 Aug 2020 10:52:17 +0000 (10:52 +0000)]
searchidx: use Perl truthiness to detect XAPIAN_FLUSH_THRESHOLD
XAPIAN_FLUSH_THRESHOLD is a C string in the environment, so
users may be tempted to assign an empty string in in their
shell, e.g. `XAPIAN_FLUSH_THRESHOLD= <command>' instead of using
`unset' POSIX shell built-in.
With either a value of "0" or "" (empty string), Xapian will
fall back to its default (10000 documents), which causes grief
for memory-starved users.
Eric Wong [Fri, 7 Aug 2020 10:52:16 +0000 (10:52 +0000)]
index: max out XAPIAN_FLUSH_THRESHOLD if using --batch-size
If XAPIAN_FLUSH_THRESHOLD is unset, Xapian will default to
10000. That limits the effectiveness of users specifying
extremely large values of --batch-size.
While we're at it, localize the changes to globals since -index
may be eval-ed in tests (and perhaps production code in the
future).
Eric Wong [Fri, 7 Aug 2020 01:14:06 +0000 (01:14 +0000)]
index+xcpdb: rename `--no-sync' to `--no-fsync'
We'll continue supporting `--no-sync' even if its yet-to-make it
it into a release, but the term `sync' is overloaded in our
codebase which may be confusing to new hackers and users.
None of our our code nor dependencies issue the sync(2) syscall,
either, only fsync(2) and fdatasync(2).
Eric Wong [Fri, 7 Aug 2020 01:14:05 +0000 (01:14 +0000)]
index: support --xapian-only switch
This is useful for speeding up indexing runs when only Xapian
rules change but SQLite indexing doesn't change. This mostly
implies `--reindex', but does NOT pick up new messages (because
SQLite indexing needs to occur for that).
I'm leaving this undocumented in the manpage for now since it's
mainly to speed up development and testing. Users upgrading to
1.6.0 will be advised to `--reindex --rethread', anyways, due to
the threading improvements since 1.1.0-pre1.
It may make sense to document for 1.7+ when there's Xapian-only
indexing changes, though.
Eric Wong [Fri, 7 Aug 2020 01:14:04 +0000 (01:14 +0000)]
index: v2: --sequential-shard option
This gives better page cache utilization for Xapian indexing on
slow storage by improving locality for random I/O activity on
the Xapian DB.
Instead of doing a single-pass to index both SQLite and Xapian;
this indexes them separately. The first pass is identical to
indexlevel=basic: it indexes both over.sqlite3 and msgmap.sqlite3.
Subsequent passes only operate on a single Xapian shard for
documents belonging to that shard. Given enough shards, each
individual shard can be made small enough to fit into the kernel
page cache and avoid HDD seeks for read activity.
Doing rough tests with a busy system with a 7200 RPM HDD with ext4,
full indexing of LKML (9 epochs) goes from ~80 hours (-j0) to
~30 hours (-j8) with 16GB RAM with 7 shards configured and fsync(2)
disabled (--no-sync) and `--batch-size=10m'.
Eric Wong [Fri, 7 Aug 2020 01:14:03 +0000 (01:14 +0000)]
v2writable: fix rethread cleanup
We need to drop old ghosts properly while inside the
transaction, otherwise it becomes a no-op. This isn't a big
deal, as it only results in a few dangling DB rows and a
small amount of wasted space.
Eric Wong [Fri, 7 Aug 2020 01:14:02 +0000 (01:14 +0000)]
xapcmd: drop outdated comment
We replaced Xtmpdir with File::Temp->newdir in
commit 2a3e3a0469f54f6a4f80bf04614e5ddd794a6c5e
("xapcmd: replace Xtmpdirs with File::Temp->newdir")
but forgot to remove the outdated comment.
Eric Wong [Fri, 7 Aug 2020 01:14:01 +0000 (01:14 +0000)]
xapcmd: remove redundant searchidx require
We already "use" it starting with commit cd8dd7b08fddc7c2b5f218c3fcaa5dca5f9ad945
("search: support SWIG-generated Xapian.pm"),
so there's no need to require it redundantly.
Eric Wong [Fri, 7 Aug 2020 01:14:00 +0000 (01:14 +0000)]
xapcmd: quietly no-op on indexlevel=basic
I find myself mindlessly adding "-c" to public-inbox-index,
and other users may do the same. Instead of erroring out,
we'll just silently ignore it, for now and allow
public-inbox-compact to work on SQLite-only inboxes.
We'll only check for xapian-compact if search exists, since
it won't be needed in case we support SQLite VACUUM.
Eric Wong [Fri, 7 Aug 2020 10:15:04 +0000 (10:15 +0000)]
syscall: support sparc64 (and maybe other big-endian systems)
Thanks to the GCC compile farm project, we can wire up syscalls
for sparc64 and set system-specific SFD_* constants properly.
I've FINALLY figured out how to use POSIX::SigSet to generate
a usable buffer for the syscall perlfunc. This is required
for endian-neutral behavior and relevant to sparc64, at least.
There's no need for signalfd-related stuff to be constants,
either. signalfd initialization is never a hot path and a stub
subroutine for constants uses several KB of memory in the
interpreter.
We'll drop the needless SEEK_CUR import while we're importing
O_NONBLOCK, too.
Eric Wong [Thu, 6 Aug 2020 21:43:46 +0000 (21:43 +0000)]
imap: search support BODY key
This is specified in RFC 3501 but was accidentally omitted :x
I probably got it confused with TEXT, so add a comment about
TEXT being "everything" in the message.
Eric Wong [Fri, 7 Aug 2020 10:42:52 +0000 (10:42 +0000)]
www: avoid warnings on YYYYMMDD-only t= query parameter
While we always generate YYYYMMDDhhmmss query parameters
ourselves, the regexps in paginate_recent allow YYYYMMDD-only
(no hhmmss) timestamps, so don't trigger Time::Local::timegm
warnings about empty numeric comparisons on empty strings when a
client starts making up their own URLs.
Eric Wong [Thu, 6 Aug 2020 10:37:10 +0000 (10:37 +0000)]
msgmap: tmp_clone: use MEMORY journal upon reconnect
Since reindexing releases the DB handle every indexBatchSize bytes,
we need to ensure we keep the journal in-memory when reopening
the DB since this is throwaway data.
Eric Wong [Sun, 2 Aug 2020 11:22:07 +0000 (11:22 +0000)]
watch: quiet some warnings on spam mailboxes
Email::Address::XS and PublicInbox::MsgTime both emit warnings
which are likely to trigger from spam messages. Since this can
be configured to remove spam, just filter out those warnings to
avoid cluttering up stderr with useless information.
Eric Wong [Sun, 2 Aug 2020 03:09:37 +0000 (03:09 +0000)]
t/nntpd: do not fork on indexing, test v2
No need to waste resources when doing minimal work. With
PI_TEST_VERSION=2, this fixes a test failure where
Net::NNTP::DESTROY was getting called in the shard process.
We'll also get rid of an unnecessary use_ok under v2, too.
Eric Wong [Sat, 1 Aug 2020 08:12:27 +0000 (08:12 +0000)]
remove unnecessary ->header_obj calls
We used ->header_obj in the past as an optimization with
Email::MIME. That optimization is no longer necessary
with PublicInbox::Eml.
This doesn't make any functional difference even if we were to
go back to Email::MIME. However, it reduces the amount of code
we have and slightly reduces allocations with PublicInbox::Eml.
Eric Wong [Fri, 31 Jul 2020 21:36:18 +0000 (21:36 +0000)]
improve error handling on import fork / lock failures
v?fork failures seems to be the cause of locks not getting
released in -watch. Ensure lock release doesn't get skipped
in ->done for both v1 and v2 inboxes. We also need to do
everything we can to ensure DB handles, pipes and processes
get released even in the face of failure.
While we're at it, make failures around `git update-server-info'
non-fatal, since smart HTTP seems more popular anyways.
v2 changes:
- spawn: show failing command
- ensure waitpid is synchronous for inotify events
- teardown all fast-import processes on exception,
not just the failing one
- beef up lock_release error handling
- release lock on fast-import spawn failure
Eric Wong [Thu, 30 Jul 2020 23:43:50 +0000 (23:43 +0000)]
www: rework async_* to use method table
Although the ->async_next method does not take $self as
a receiver, but rather a PublicInbox::HTTP object, we may
still retrieve it to be called with the HTTP object via
UNIVERSAL->can.
Eric Wong [Wed, 29 Jul 2020 21:25:43 +0000 (21:25 +0000)]
msgmap: disable CoW for tmp_clone, too
The temporary clone starts as large as the full msgmap
and deletes will write to it randomly. So ensure it
doesn't get fragmented and slower as time goes on.
wwwlisting: fix grep call for match=domain filtering
The grep call in list_match_domain_i returns true for all inboxes,
even ones without a URL that matches the regular expression, because
the qr value passed to grep is not surrounded by slashes. Add them.
Fixes: 1988d730c0088e8b (config: support multi-value inbox.*.*url)
Eric Wong [Wed, 29 Jul 2020 08:43:09 +0000 (08:43 +0000)]
emergency: create full path to PI_EMERGENCY
It's possible for ~/.public-inbox/ to not exist if PI_CONFIG
points to an alternate location. Only noticed from the previous
patch fixing t/init.t behavior.
Eric Wong [Wed, 29 Jul 2020 11:20:00 +0000 (11:20 +0000)]
xapcmd: -xcpdb and -compact disable CoW, too
This gives an opportunity for users already suffering from CoW
fragmentation to at least get the Xapian DBs off CoW. Aside
from over.sqlite3 in v1, the SQLite DBs remain untouched; though
VACUUM support may come in the future.
Eric Wong [Tue, 28 Jul 2020 22:21:58 +0000 (22:21 +0000)]
searchidx: disable CoW for SQLite and Xapian under btrfs
SQLite and Xapian files are written randomly, thus they become
fragmented under btrfs with copy-on-write. This leads to
noticeable performance problems (and probably ENOSPC) as these
files get big.
lore/git (v2, <1GB) indexes around 20% faster with this on an
ancient SSD. lore/lkml seems to be taking forever and I'll
probably cancel it to save wear on my SSD.
Unfortunately, disabling CoW also means disabling checksumming
(and compression), so we'll be careful to only set the No_COW
attribute on regeneratable data. We want to keep CoW (and
checksums+compression) on git storage because current ref
storage is neither checksummed nor compressed, and git streams
pack output.
Eric Wong [Sat, 25 Jul 2020 21:12:21 +0000 (21:12 +0000)]
v2writable: support async git blob retrievals
This seems to speed up --reindex on smallish v2 inboxes by about
30% on both HDD and SSD. lore/git (~1GB) on an SSD even gives a
30% improvement with 3 shards. I'm only seeing a ~4% speedup on
LKML with a SATA SSD (which is difficult to repeat because it
takes around 4 hours).
Testing LKML on an HDD will take much more time...
Eric Wong [Sun, 26 Jul 2020 06:57:30 +0000 (06:57 +0000)]
imap: introduce and use Git->async_prefetch
We can keep the git process more active by sending another
request to it while fetch_run_ops() is running. This
parallelization speeds up mutt's initial FETCH for headers by
around ~35%(!).
Eric Wong [Sun, 26 Jul 2020 09:31:10 +0000 (09:31 +0000)]
index: --compact respects --jobs
And -compact supports --jobs=0 like -index to disable parallel
execution. Running three xapian-compact processes in parallel
on a USB 2.0 HDD is pretty painful.
Eric Wong [Sun, 26 Jul 2020 19:43:06 +0000 (19:43 +0000)]
overidx: fix compatibility with current versions
We still need to use SQL_BLOB to ensure existing versions of
public-inbox can read over.sqlite3 because they're still using
{sqlite_unicode}. This partially reverts commit e9fc1290ead44e06d20ff58e0a6acb5306d4fbe2.
Eric Wong [Fri, 24 Jul 2020 05:56:01 +0000 (05:56 +0000)]
searchidx: make v1 indexing closer to v2
We'll switch to using IdxStack here to ensure we get repeatable
results and ascending THREADIDs according to git chronology.
This means we'll need a two-pass reindex to index existing
messages before indexing new messages.
Since we no longer have a long-lived git-log process, we don't
have to worry about old Xapian referencing the git-log pipe
w/o FD_CLOEXEC, either.
Eric Wong [Fri, 24 Jul 2020 05:56:00 +0000 (05:56 +0000)]
searchidx: rename _xdb_{acquire,release} => idx_
The "xdb" prefix was inaccurate since it's used by
indexlevel=basic, which is Xapian-free. The '_' (underscore)
prefix was also wrong for a method which is called across
package boundaries.
Eric Wong [Fri, 24 Jul 2020 05:55:57 +0000 (05:55 +0000)]
v2writable: get rid of {reindex_pipe} field
Since normal per-epoch indexing no longer holds a "git log"
process open, we don't need to worry about not sharing the
pipe with forked shards when we restart the indexer.
While we're in the area, better describe what `unindex' does,
since it's a rarely-used but necessary code path.