about summary refs log tree commit homepage
path: root/lib/PublicInbox/SearchIdx.pm
diff options
context:
space:
mode:
authorEric Wong (Contractor, The Linux Foundation) <e@80x24.org>2018-02-22 01:49:08 +0000
committerEric Wong (Contractor, The Linux Foundation) <e@80x24.org>2018-02-22 18:33:46 +0000
commit9ecbfc09928dada28094fd3fc79e91a5472b27ea (patch)
treea829ab7765f45e139e8a9d5de1c3784fc26bbf69 /lib/PublicInbox/SearchIdx.pm
parenta81ad9c4b1b5d8c2ae8444b6dcb8710bd361f628 (diff)
downloadpublic-inbox-9ecbfc09928dada28094fd3fc79e91a5472b27ea.tar.gz
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
Diffstat (limited to 'lib/PublicInbox/SearchIdx.pm')
-rw-r--r--lib/PublicInbox/SearchIdx.pm80
1 files changed, 58 insertions, 22 deletions
diff --git a/lib/PublicInbox/SearchIdx.pm b/lib/PublicInbox/SearchIdx.pm
index c6c5bd25..cc7e7ec9 100644
--- a/lib/PublicInbox/SearchIdx.pm
+++ b/lib/PublicInbox/SearchIdx.pm
@@ -51,7 +51,7 @@ sub git_unquote ($) {
 }
 
 sub new {
-        my ($class, $ibx, $creat) = @_;
+        my ($class, $ibx, $creat, $part) = @_;
         my $mainrepo = $ibx; # for "public-inbox-index" w/o entry in config
         my $git_dir = $mainrepo;
         my ($altid, $git);
@@ -83,7 +83,10 @@ sub new {
         if ($version == 1) {
                 $self->{lock_path} = "$mainrepo/ssoma.lock";
         } elsif ($version == 2) {
-                $self->{lock_path} = "$mainrepo/inbox.lock";
+                defined $part or die "partition is required for v2\n";
+                # partition is a number or "all"
+                $self->{partition} = $part;
+                $self->{lock_path} = undef;
                 $self->{msgmap_path} = "$mainrepo/msgmap.sqlite3";
         } else {
                 die "unsupported inbox version=$version\n";
@@ -119,14 +122,16 @@ sub _xdb_acquire {
 sub _lock_acquire {
         my ($self) = @_;
         croak 'already locked' if $self->{lockfh};
-        sysopen(my $lockfh, $self->{lock_path}, O_WRONLY|O_CREAT) or
-                die "failed to open lock $self->{lock_path}: $!\n";
+        my $lock_path = $self->{lock_path} or return;
+        sysopen(my $lockfh, $lock_path, O_WRONLY|O_CREAT) or
+                die "failed to open lock $lock_path: $!\n";
         flock($lockfh, LOCK_EX) or die "lock failed: $!\n";
         $self->{lockfh} = $lockfh;
 }
 
 sub _lock_release {
         my ($self) = @_;
+        return unless $self->{lock_path};
         my $lockfh = delete $self->{lockfh} or croak 'not locked';
         flock($lockfh, LOCK_UN) or die "unlock failed: $!\n";
         close $lockfh or die "close failed: $!\n";
@@ -138,8 +143,8 @@ sub add_val ($$$) {
         $doc->add_value($col, $num);
 }
 
-sub add_values ($$$) {
-        my ($smsg, $bytes, $num) = @_;
+sub add_values ($$$$) {
+        my ($smsg, $bytes, $num, $lines) = @_;
 
         my $ts = $smsg->ts;
         my $doc = $smsg->{doc};
@@ -149,8 +154,7 @@ sub add_values ($$$) {
 
         defined($bytes) and add_val($doc, &PublicInbox::Search::BYTES, $bytes);
 
-        add_val($doc, &PublicInbox::Search::LINES,
-                        $smsg->{mime}->body_raw =~ tr!\n!\n!);
+        add_val($doc, &PublicInbox::Search::LINES, $lines);
 
         my $yyyymmdd = strftime('%Y%m%d', gmtime($ts));
         add_val($doc, PublicInbox::Search::YYYYMMDD, $yyyymmdd);
@@ -281,6 +285,7 @@ sub add_message {
 
         my ($doc_id, $old_tid);
         my $mid = mid_clean(mid_mime($mime));
+        my $threader = $self->{threader};
 
         eval {
                 die 'Message-ID too long' if length($mid) > MAX_MID_SIZE;
@@ -289,19 +294,22 @@ sub add_message {
                         # convert a ghost to a regular message
                         # it will also clobber any existing regular message
                         $doc_id = $smsg->{doc_id};
-                        $old_tid = $smsg->thread_id;
+                        $old_tid = $smsg->thread_id unless $threader;
                 }
                 $smsg = PublicInbox::SearchMsg->new($mime);
                 my $doc = $smsg->{doc};
                 $doc->add_term('XMID' . $mid);
 
                 my $subj = $smsg->subject;
+                my $xpath;
                 if ($subj ne '') {
-                        my $path = $self->subject_path($subj);
-                        $doc->add_term('XPATH' . id_compress($path));
+                        $xpath = $self->subject_path($subj);
+                        $xpath = id_compress($xpath);
+                        $doc->add_term('XPATH' . $xpath);
                 }
 
-                add_values($smsg, $bytes, $num);
+                my $lines = $mime->body_raw =~ tr!\n!\n!;
+                add_values($smsg, $bytes, $num, $lines);
 
                 my $tg = $self->term_generator;
 
@@ -350,9 +358,16 @@ sub add_message {
                         index_body($tg, \@orig, $doc) if @orig;
                 });
 
-                link_message($self, $smsg, $old_tid);
+                # populates smsg->references for smsg->to_doc_data
+                my $refs = parse_references($smsg);
+                my $data = $smsg->to_doc_data($blob);
+                if ($threader) {
+                        $threader->thread_msg($mid, $smsg->ts, $xpath, $data);
+                } else {
+                        link_message($self, $smsg, $refs, $old_tid);
+                }
                 $tg->index_text($mid, 1, 'XM');
-                $doc->set_data($smsg->to_doc_data($blob));
+                $doc->set_data($data);
 
                 if (my $altid = $self->{-altid}) {
                         foreach my $alt (@$altid) {
@@ -424,8 +439,8 @@ sub next_thread_id {
         $last_thread_id;
 }
 
-sub link_message {
-        my ($self, $smsg, $old_tid) = @_;
+sub parse_references ($) {
+        my ($smsg) = @_;
         my $doc = $smsg->{doc};
         my $mid = $smsg->mid;
         my $mime = $smsg->{mime};
@@ -436,7 +451,6 @@ sub link_message {
         my @refs = (($hdr->header_raw('References') || '') =~ /<([^>]+)>/g);
         push(@refs, (($hdr->header_raw('In-Reply-To') || '') =~ /<([^>]+)>/g));
 
-        my $tid;
         if (@refs) {
                 my %uniq = ($mid => 1);
                 my @orig_refs = @refs;
@@ -452,25 +466,31 @@ sub link_message {
                         push @refs, $ref;
                 }
         }
+        $smsg->{references} = '<'.join('> <', @refs).'>' if @refs;
+        \@refs
+}
 
-        if (@refs) {
-                $smsg->{references} = '<'.join('> <', @refs).'>';
+sub link_message {
+        my ($self, $smsg, $refs, $old_tid) = @_;
+        my $tid;
+
+        if (@$refs) {
 
                 # first ref *should* be the thread root,
                 # but we can never trust clients to do the right thing
-                my $ref = shift @refs;
+                my $ref = shift @$refs;
                 $tid = $self->_resolve_mid_to_tid($ref);
                 $self->merge_threads($tid, $old_tid) if defined $old_tid;
 
                 # the rest of the refs should point to this tid:
-                foreach $ref (@refs) {
+                foreach $ref (@$refs) {
                         my $ptid = $self->_resolve_mid_to_tid($ref);
                         merge_threads($self, $tid, $ptid);
                 }
         } else {
                 $tid = defined $old_tid ? $old_tid : $self->next_thread_id;
         }
-        $doc->add_term('G' . $tid);
+        $smsg->{doc}->add_term('G' . $tid);
 }
 
 sub index_blob {
@@ -798,4 +818,20 @@ sub DESTROY {
         $_[0]->{lockfh} = undef;
 }
 
+# remote_* subs are only used by SearchIdxPart and SearchIdxThread:
+sub remote_commit {
+        my ($self) = @_;
+        print { $self->{w} } "commit\n" or die "failed to write commit: $!";
+}
+
+sub remote_close {
+        my ($self) = @_;
+        my $pid = delete $self->{pid} or die "no process to wait on\n";
+        my $w = delete $self->{w} or die "no pipe to write to\n";
+        print $w "close\n" or die "failed to write to pid:$pid: $!\n";
+        close $w or die "failed to close pipe for pid:$pid: $!\n";
+        waitpid($pid, 0) == $pid or die "remote process did not finish";
+        $? == 0 or die ref($self)." exited with: $?";
+}
+
 1;