Replicated documents do not properly go through validate_doc_update functions
-----------------------------------------------------------------------------
Key: COUCHDB-915
URL: https://issues.apache.org/jira/browse/COUCHDB-915
Project: CouchDB
Issue Type: Bug
Components: Replication
Affects Versions: 1.0.1
Environment: Linux RedHat AS 4 x86
Reporter: Oguzhan Eris
Before a replicated document gets written, it undergoes a validation check just
like a normal doc update would, but in the case of replication, the "oldDoc"
variable for a validate_doc_update function does not represent the latest
revision of the document.
Imagine two couchdb instances each with the same starting doc.
{_id:"docA", lastModifiedTime:1}
now nodeA updates the doc to
{_id:"docA", lastModifiedTime:2}
and nodeB updates independently to {_id:"docA",lastModifiedTime:30}
and imagine both nodes already having a validate_doc_update function that says
(if oldDoc && Number(oldDoc.lastModifiedTime) >=
Number(newDoc.lastModifiedTime)) { throw ([forbidden, "already have a more
recent doc"]);
so each doc has properly gone through a validation check to get their second
revision, and when we replicate from nodeB to nodeA, we should expect that the
document should indeed be updated to lastModifiedTime:30 and when we replicate
from nodeA to nodeB we should NOT get lastModifiedTime:2 since our validation
function should prevent it.
What happens however is that when nodeA replicates to nodeB, nodeB's validation
function gets called with the Rev1 as the oldDoc argument, and in this
situation lastModifiedTime:2 does get the "ok" from the validate function since
it's comparing against the first revision with lastModifiedTime:1 instead of
latest revision lastModifiedTime:30
This happens in couch_db.erl
prep_and_validate_replicated_updates(Db, [Bucket|RestBuckets],
[OldInfo|RestOldInfo], AccPrepped, AccErrors)
more specifically:
lists:foldl(
fun(#doc{id=Id,revs={Pos, [RevId|_]}}=Doc, {AccValidated,
AccErrors2}) ->
case dict:find({Pos, RevId}, LeafRevsFullDict) of
{ok, {Start, Path}} ->
% our unflushed doc is a leaf node. Go back on the path
% to find the previous rev that's on disk.
I am not sure what the reasoning behind this is, but to me, it'd make more
sense to compare (or at least have a way to compare) the latest revision
locally as the oldDoc in a validate_doc_update function.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
