[
https://issues.apache.org/jira/browse/COUCHDB-864?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12900809#action_12900809
]
Robert Newson commented on COUCHDB-864:
---------------------------------------
JIRA is probably not the easiest way to convey the gist of the discussion Adam
and I had today.
The first thing to clarify is that there's no bug as such, just a (significant)
inefficiency. Mochiweb closes the request and the http client works correctly
when that happens.
I discovered the problem as I'm doing performance testing. I'm using nodeload
(a node.js-based http benchmarking kit). I tell it to keep the connections
alive. This works fine in my original test where I upload standalone
attachments. When I switched to uploading multipart/related, so that I can set
json metadata as well, it blew up with 'connection reset by peer'. Assuming, as
one does, that it was a node.js bug, I tried the same upload in other ways and
they all behaved the same way.
CouchDB should be able to keep the connections alive for multiple requests in
most cases. It does so, as far as I can tell, in every case exception
multipart/related PUT's. Having read the related code, and discussed it with
Adam on IRC, it's clear why.
I'll post a further patch later on when I've had more time to tidy and test.
> multipart/related PUT's always close the connection.
> ----------------------------------------------------
>
> Key: COUCHDB-864
> URL: https://issues.apache.org/jira/browse/COUCHDB-864
> Project: CouchDB
> Issue Type: Bug
> Components: Database Core
> Reporter: Robert Newson
>
> I noticed that mochiweb always closes the connection when doing a
> multipart/related PUT (to insert the JSON document and accompanying
> attachments in one call). Ultimately it's because we call recv(0) and not
> recv_body, thus consuming more data than we actually process. Mochiweb
> notices that there is unread data on the socket and closes the connection.
> This impacts replication with attachments, as I believe they go through this
> code path (and, thus, are forever reconnecting).
> The code below demonstrates a fix for this issue but isn't good enough for
> trunk. Adam provided the important process dictionary fix.
> ---
> src/couchdb/couch_doc.erl | 1 +
> src/couchdb/couch_httpd_db.erl | 13 +++++++++----
> 2 files changed, 10 insertions(+), 4 deletions(-)
> diff --git a/src/couchdb/couch_doc.erl b/src/couchdb/couch_doc.erl
> index 5009f8f..f8c874b 100644
> --- a/src/couchdb/couch_doc.erl
> +++ b/src/couchdb/couch_doc.erl
> @@ -455,6 +455,7 @@ doc_from_multi_part_stream(ContentType, DataFun) ->
> Parser ! {get_doc_bytes, self()},
> receive
> {doc_bytes, DocBytes} ->
> + erlang:put(mochiweb_request_recv, true),
> Doc = from_json_obj(?JSON_DECODE(DocBytes)),
> % go through the attachments looking for 'follows' in the data,
> % replace with function that reads the data from MIME stream.
> diff --git a/src/couchdb/couch_httpd_db.erl b/src/couchdb/couch_httpd_db.erl
> index b0fbe8d..eff7d67 100644
> --- a/src/couchdb/couch_httpd_db.erl
> +++ b/src/couchdb/couch_httpd_db.erl
> @@ -651,12 +651,13 @@ db_doc_req(#httpd{method='PUT'}=Req, Db, DocId) ->
> } = parse_doc_query(Req),
> couch_doc:validate_docid(DocId),
>
> + Len = couch_httpd:header_value(Req,"Content-Length"),
> Loc = absolute_uri(Req, "/" ++ ?b2l(Db#db.name) ++ "/" ++ ?b2l(DocId)),
> RespHeaders = [{"Location", Loc}],
> case couch_util:to_list(couch_httpd:header_value(Req, "Content-Type")) of
> ("multipart/related;" ++ _) = ContentType ->
> {ok, Doc0} = couch_doc:doc_from_multi_part_stream(ContentType,
> - fun() -> receive_request_data(Req) end),
> + fun() -> receive_request_data(Req, Len) end),
> Doc = couch_doc_from_req(Req, DocId, Doc0),
> update_doc(Req, Db, DocId, Doc, RespHeaders, UpdateType);
> _Else ->
> @@ -775,9 +776,13 @@ send_docs_multipart(Req, Results, Options) ->
> couch_httpd:send_chunk(Resp, <<"--">>),
> couch_httpd:last_chunk(Resp).
>
> -receive_request_data(Req) ->
> - {couch_httpd:recv(Req, 0), fun() -> receive_request_data(Req) end}.
> -
> +receive_request_data(Req, undefined) ->
> + receive_request_data(Req, 0);
> +receive_request_data(Req, Len) when is_list(Len)->
> + Remaining = list_to_integer(Len),
> + Bin = couch_httpd:recv(Req, Remaining),
> + {Bin, fun() -> receive_request_data(Req, Remaining - iolist_size(Bin))
> end}.
> +
> update_doc_result_to_json({{Id, Rev}, Error}) ->
> {_Code, Err, Msg} = couch_httpd:error_info(Error),
> {[{id, Id}, {rev, couch_doc:rev_to_str(Rev)},
> --
> 1.7.2.2
> Umbra
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
