Up to 1.0.1, it was possible to loose the identity length of an attachment that was compressed by CouchDB. The only case known where this happened was after doing a local to local replication. This was found and fixed recently in 1.0.x with a patch from Juuso Väänänen,
However, existing databases created with versions up to 1.0.1, have the attachment identify length lost forever. For this cases, the multipart/related or multipart/mixed APIs used to GET a document, return attachments stubs with a "length" field matching the length of the attachment in compressed form, and the part of the multipart stream that contains the attachment, contains it in uncompressed form.
This poses a serious problem to the CouchDB multipart parser, an essential part used by the new replicator that's likely to land on the 1.2 release. Without this fix, it may not be possible to do pull replications, with the new replicator, from versions 0.11.0, 0.11.1, 0.11.2, 1.0.0 and 1.0.1 ONLY when the source database was the product of a local to local replication (and there are attachments stored in compressed form, as described by
The following patch fixes this issue by always sending the attachments in compressed form. It's meant to be applied against branches 0.11.x, 1.0.x, 1.1.x and trunk.