nickva commented on a change in pull request #806: FEEDBACK ONLY: Compactor
optimize emsort
URL: https://github.com/apache/couchdb/pull/806#discussion_r139160022
##########
File path: src/couch/src/couch_db_updater.erl
##########
@@ -1214,103 +1265,141 @@ copy_compact(Db, NewDb0, Retry) ->
NewDb4 = NewDb3
end,
- commit_compaction_data(NewDb4#db{update_seq=Db#db.update_seq}).
-
-
-start_copy_compact(#db{}=Db) ->
- erlang:put(io_priority, {db_compact, Db#db.name}),
- #db{name=Name, filepath=Filepath, options=Options, header=Header} = Db,
- couch_log:debug("Compaction process spawned for db \"~s\"", [Name]),
-
- {ok, NewDb, DName, DFd, MFd, Retry} =
- open_compaction_files(Name, Header, Filepath, Options),
- erlang:monitor(process, MFd),
-
- % This is a bit worrisome. init_db/4 will monitor the data fd
- % but it doesn't know about the meta fd. For now I'll maintain
- % that the data fd is the old normal fd and meta fd is special
- % and hope everything works out for the best.
- unlink(DFd),
-
- NewDb1 = copy_purge_info(Db, NewDb),
- NewDb2 = copy_compact(Db, NewDb1, Retry),
- NewDb3 = sort_meta_data(NewDb2),
- NewDb4 = commit_compaction_data(NewDb3),
- NewDb5 = copy_meta_data(NewDb4),
- NewDb6 = sync_header(NewDb5, db_to_header(NewDb5, NewDb5#db.header)),
- close_db(NewDb6),
-
- ok = couch_file:close(MFd),
- gen_server:cast(Db#db.main_pid, {compact_done, DName}).
+ CompSt#comp_st{
+ new_db = NewDb4#db{
+ update_seq = Db#db.update_seq
+ }
+ }.
-open_compaction_files(DbName, SrcHdr, DbFilePath, Options) ->
- DataFile = DbFilePath ++ ".compact.data",
- MetaFile = DbFilePath ++ ".compact.meta",
- {ok, DataFd, DataHdr} = open_compaction_file(DataFile),
- {ok, MetaFd, MetaHdr} = open_compaction_file(MetaFile),
- DataHdrIsDbHdr = couch_db_header:is_header(DataHdr),
- case {DataHdr, MetaHdr} of
- {#comp_header{}=A, #comp_header{}=A} ->
- DbHeader = A#comp_header.db_header,
- Db0 = init_db(DbName, DataFile, DataFd, DbHeader, Options),
- Db1 = bind_emsort(Db0, MetaFd, A#comp_header.meta_state),
- {ok, Db1, DataFile, DataFd, MetaFd, Db0#db.id_tree};
- _ when DataHdrIsDbHdr ->
- ok = reset_compaction_file(MetaFd, couch_db_header:from(SrcHdr)),
- Db0 = init_db(DbName, DataFile, DataFd, DataHdr, Options),
- Db1 = bind_emsort(Db0, MetaFd, nil),
- {ok, Db1, DataFile, DataFd, MetaFd, Db0#db.id_tree};
- _ ->
- Header = couch_db_header:from(SrcHdr),
- ok = reset_compaction_file(DataFd, Header),
- ok = reset_compaction_file(MetaFd, Header),
- Db0 = init_db(DbName, DataFile, DataFd, Header, Options),
- Db1 = bind_emsort(Db0, MetaFd, nil),
- {ok, Db1, DataFile, DataFd, MetaFd, nil}
- end.
+copy_docs(Db, #db{fd = DestFd} = NewDb, MixedInfos, Retry) ->
+ DocInfoIds = [Id || #doc_info{id=Id} <- MixedInfos],
+ LookupResults = couch_btree:lookup(Db#db.id_tree, DocInfoIds),
+ % COUCHDB-968, make sure we prune duplicates during compaction
+ NewInfos0 = lists:usort(fun(#full_doc_info{id=A}, #full_doc_info{id=B}) ->
+ A =< B
+ end, merge_lookups(MixedInfos, LookupResults)),
+ NewInfos1 = lists:map(fun(Info) ->
+ {NewRevTree, FinalAcc} = couch_key_tree:mapfold(fun
+ (_Rev, #leaf{ptr=Sp}=Leaf, leaf, SizesAcc) ->
+ {Body, AttInfos} = copy_doc_attachments(Db, Sp, DestFd),
+ % In the future, we should figure out how to do this for
+ % upgrade purposes.
+ EJsonBody = case is_binary(Body) of
+ true ->
+ couch_compress:decompress(Body);
+ false ->
+ Body
+ end,
+ SummaryChunk = make_doc_summary(NewDb, {Body, AttInfos}),
+ ExternalSize = ?term_size(EJsonBody),
+ {ok, Pos, SummarySize} = couch_file:append_raw_chunk(
+ DestFd, SummaryChunk),
+ AttSizes = [{element(3,A), element(4,A)} || A <- AttInfos],
+ NewLeaf = Leaf#leaf{
+ ptr = Pos,
+ sizes = #size_info{
+ active = SummarySize,
+ external = ExternalSize
+ },
+ atts = AttSizes
+ },
+ {NewLeaf, add_sizes(leaf, NewLeaf, SizesAcc)};
+ (_Rev, _Leaf, branch, SizesAcc) ->
+ {?REV_MISSING, SizesAcc}
+ end, {0, 0, []}, Info#full_doc_info.rev_tree),
+ {FinalAS, FinalES, FinalAtts} = FinalAcc,
+ TotalAttSize = lists:foldl(fun({_, S}, A) -> S + A end, 0, FinalAtts),
+ NewActiveSize = FinalAS + TotalAttSize,
+ NewExternalSize = FinalES + TotalAttSize,
+ ?COMP_EVENT(seq_copy),
+ Info#full_doc_info{
+ rev_tree = NewRevTree,
+ sizes = #size_info{
+ active = NewActiveSize,
+ external = NewExternalSize
+ }
+ }
+ end, NewInfos0),
-open_compaction_file(FilePath) ->
- case couch_file:open(FilePath, [nologifmissing]) of
- {ok, Fd} ->
- case couch_file:read_header(Fd) of
- {ok, Header} -> {ok, Fd, Header};
- no_valid_header -> {ok, Fd, nil}
- end;
- {error, enoent} ->
- {ok, Fd} = couch_file:open(FilePath, [create]),
- {ok, Fd, nil}
- end.
+ NewInfos = stem_full_doc_infos(Db, NewInfos1),
+ RemoveSeqs =
+ case Retry of
+ nil ->
+ [];
+ OldDocIdTree ->
+ % Compaction is being rerun to catch up to writes during the
+ % first pass. This means we may have docs that already exist
+ % in the seq_tree in the .data file. Here we lookup any old
+ % update_seqs so that they can be removed.
+ Ids = [Id || #full_doc_info{id=Id} <- NewInfos],
+ Existing = couch_btree:lookup(OldDocIdTree, Ids),
+ [Seq || {ok, #full_doc_info{update_seq=Seq}} <- Existing]
+ end,
+ {ok, SeqTree} = couch_btree:add_remove(
+ NewDb#db.seq_tree, NewInfos, RemoveSeqs),
-reset_compaction_file(Fd, Header) ->
- ok = couch_file:truncate(Fd, 0),
- ok = couch_file:write_header(Fd, Header).
+ EMSortFd = couch_emsort:get_fd(NewDb#db.id_tree),
+ {ok, LocSizes} = couch_file:append_terms(EMSortFd, NewInfos),
Review comment:
Nice! Good idea to create `append_terms` and use it here.
This will use the default (`snappy`) compression. Wondering if we should try
to benchmark with no compression here. Snappy will incur an overhead (even if
it doesn't have much to do, just jumping into a NIF), and in `fprof` profiling
saw it eating a decent chunk of time. So try to pass the compression explicitly
here or maybe just change `couch_file:append_terms/2` to not use snappy as a
default compression but make it `none` if user doesn't pass it in.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
