Ensure doc groups are sorted before merging them2735-duplicate-docs

We had been implicitly assuming that clients send us sorted groups, but
unsurprisingly that's not always the case. The additional sorting here
should be redundant, but the consequences of merging unsorted groups are
severe -- we can end up with uniqueness violations on the primary key in
the database -- and so we add an additional sort here.

COUCHDB-2735

1 files changed, 11 insertions, 3 deletions

diff --git a/src/couchdb/couch_db_updater.erl b/src/couchdb/couch_db_updater.erl
index 947669cb1..c92097f62 100644
--- a/src/couchdb/couch_db_updater.erl
+++ b/src/couchdb/couch_db_updater.erl
@@ -222,7 +222,7 @@ handle_cast(Msg, #db{name = Name} = Db) ->
 
 handle_info({update_docs, Client, GroupedDocs, NonRepDocs, MergeConflicts,
         FullCommit}, Db) ->
-    GroupedDocs2 = [[{Client, D} || D <- DocGroup] || DocGroup <- GroupedDocs],
+    GroupedDocs2 = sort_and_tag_groups(Client, GroupedDocs),
     if NonRepDocs == [] ->
         {GroupedDocs3, Clients, FullCommit2} = collect_updates(GroupedDocs2,
                 [Client], MergeConflicts, FullCommit);
@@ -291,8 +291,7 @@ collect_updates(GroupedDocsAcc, ClientsAcc, MergeConflicts, FullCommit) ->
         % updaters than deal with their possible conflicts, and local docs
         % writes are relatively rare. Can be optmized later if really needed.
         {update_docs, Client, GroupedDocs, [], MergeConflicts, FullCommit2} ->
-            GroupedDocs2 = [[{Client, Doc} || Doc <- DocGroup]
-                    || DocGroup <- GroupedDocs],
+            GroupedDocs2 = sort_and_tag_groups(Client, GroupedDocs),
             GroupedDocsAcc2 =
                 merge_updates(GroupedDocsAcc, GroupedDocs2, []),
             collect_updates(GroupedDocsAcc2, [Client | ClientsAcc],
@@ -302,6 +301,15 @@ collect_updates(GroupedDocsAcc, ClientsAcc, MergeConflicts, FullCommit) ->
     end.
 
 
+sort_and_tag_groups(Client, GroupedDocs) ->
+    % These groups should already be sorted but sometimes clients misbehave.
+    % The merge_updates function will fail and the database can end up with
+    % duplicate documents if the incoming groups are not sorted, so as a sanity
+    % check we sort them again here. See COUCHDB-2735.
+    Cmp = fun([{#doc{id=A}, _}|_], [{#doc{id=B}, _}|_]) -> A < B end,
+    SortedGroups = lists:sort(Cmp, GroupedDocs),
+    [[{Client, D} || D <- DocGroup] || DocGroup <- SortedGroups].
+
 btree_by_seq_split(#doc_info{id=Id, high_seq=KeySeq, revs=Revs}) ->
     {RevInfos, DeletedRevInfos} = lists:foldl(
         fun(#rev_info{deleted = false, seq = Seq} = Ri, {Acc, AccDel}) ->