]> Sergey Matveev's repositories - public-inbox.git/commitdiff
extsearchidx: checkpoint releases locks
authorEric Wong <e@80x24.org>
Tue, 15 Dec 2020 02:02:21 +0000 (02:02 +0000)
committerEric Wong <e@80x24.org>
Thu, 17 Dec 2020 19:13:12 +0000 (19:13 +0000)
--reindex can take many hours or days, ensure we release
locks according to --batch-size so automated fetch+index
jobs can write new data to indices while we update old data.

lib/PublicInbox/ExtSearchIdx.pm
lib/PublicInbox/V2Writable.pm

index f29a84e3d1d3fa6db23779e739cc5da2ddb2cd26..3b021a1bc212d8aa25a38be929114d4be25beddf 100644 (file)
@@ -404,7 +404,7 @@ sub _reindex_finalize ($$$) {
        my $orig_smsg = $req->{orig_smsg} // die 'BUG: no {orig_smsg}';
        my $docid = $smsg->{num} = $orig_smsg->{num};
        $self->{oidx}->add_overview($eml, $smsg); # may rethread
-       $self->{transact_bytes} += $smsg->{bytes};
+       check_batch_limit({ %$sync, new_smsg => $smsg });
        if ($nr == 1) { # likely, all good
                $self->idx_shard($docid)->shard_reindex_docid($docid);
                return;
@@ -537,18 +537,8 @@ sub eidxq_process ($$) { # for reindexing
                        warn "E: #$docid does not exist in over\n";
                }
                $del->execute($docid);
+               ++${$sync->{nr}};
 
-               my $cur = ++${$sync->{nr}};
-
-               # shards flush on their own, just don't queue up too many
-               # deletes
-               if ($self->{transact_bytes} >= $self->{batch_bytes}) {
-                       $self->git->async_wait_all;
-                       $self->{oidx}->commit_lazy;
-                       $self->{oidx}->begin_lazy;
-                       $pr->("reindexed $cur/$tot\n") if $pr;
-                       $self->{transact_bytes} = 0;
-               }
                # this is only for SIGUSR1, shards do their own accounting:
                reindex_checkpoint($self, $sync) if ${$sync->{need_checkpoint}};
        }
@@ -709,7 +699,10 @@ sub eidx_sync { # main entry point
        for my $ibx (@{$self->{ibx_list}}) {
                $ibx->{-ibx_id} //= $self->{oidx}->ibx_id($ibx->eidx_key);
        }
-       eidx_reindex($self, $sync) if delete($opt->{reindex});
+       if (delete($opt->{reindex})) {
+               $sync->{checkpoint_unlocks} = 1;
+               eidx_reindex($self, $sync);
+       }
 
        # don't use $_ here, it'll get clobbered by reindex_checkpoint
        for my $ibx (@{$self->{ibx_list}}) {
index 572eb4184bafc2a86e1c787b12e3855bc1a799a5..97dbf328a3dde7b503b91cbe3d9e785e0437ae2c 100644 (file)
@@ -884,7 +884,7 @@ sub reindex_checkpoint ($$) {
        my $mm_tmp = $sync->{mm_tmp};
        $mm_tmp->atfork_prepare if $mm_tmp;
        die 'BUG: {im} during reindex' if $self->{im};
-       if ($self->{ibx_map}) {
+       if ($self->{ibx_map} && !$sync->{checkpoint_unlocks}) {
                checkpoint($self, 1); # no need to release lock on pure index
        } else {
                $self->done; # release lock