blackwell-systems opened a new issue, #6162: URL: https://github.com/apache/couchdb/issues/6162
### Version 3.5.3 ### Describe the problem you're encountering Two target databases that get the same replications, in the same order, from the same sources can end up with a different state for a document: one ends deleted, the other ends live with an old revision as the winner. The only difference between the targets is a design doc whose `validate_doc_update` accepts everything. It shows up when a target first receives a revision with a short history (from a replica with a low `revs_limit`) and later receives the same revision again with its full history. ### Expected Behaviour I would expect replication to give the same result whether or not the target has a `validate_doc_update` function that accepts the document. ### Steps to Reproduce Needs `curl` and `jq`. Single node, one-shot `_replicate`, run in order. ```sh COUCH=http://adm:[email protected]:5984 J='Content-Type: application/json' for db in a c b d t1 t2; do curl -sS -X PUT "$COUCH/$db" > /dev/null; done curl -sS -X PUT "$COUCH/b/_revs_limit" -d 1 > /dev/null curl -sS -X PUT "$COUCH/d/_revs_limit" -d 2 > /dev/null curl -sS -X PUT "$COUCH/t2/_design/accept" -H "$J" \ -d '{"validate_doc_update":"function(newDoc, oldDoc, userCtx, secObj) {}"}' > /dev/null rep() { curl -sS -X POST "$COUCH/_replicate" -H "$J" \ -d "{\"source\":\"$COUCH/$1\",\"target\":\"$COUCH/$2\"${3:+,$3}}" > /dev/null; } ONLY_X='"doc_ids":["X"]' # create X (1-a) on a, replicate to c REV=$(curl -sS -X PUT "$COUCH/a/X" -H "$J" -d '{"v":"a"}' | jq -r .rev) rep a c "$ONLY_X" # 6 unrelated docs on a curl -sS -X POST "$COUCH/a/_bulk_docs" -H "$J" -d '{"docs":[ {"_id":"f1"},{"_id":"f2"},{"_id":"f3"},{"_id":"f4"},{"_id":"f5"},{"_id":"f6"}]}' > /dev/null # edit X (2-b) on a, replicate to b (revs_limit 1, so b keeps 2-b with no ancestors) REV=$(curl -sS -X PUT "$COUCH/a/X" -H "$J" -d "{\"_rev\":\"$REV\",\"v\":\"b\"}" | jq -r .rev) rep a b "$ONLY_X" # both targets: c then b, giving two separate roots 1-a and 2-b for t in t1 t2; do rep c $t; rep b $t; done # both targets: replicate a with one worker, batch size 1 for t in t1 t2; do rep a $t '"worker_processes":1,"worker_batch_size":1'; done # delete X (3-d) on a, replicate to d (revs_limit 2), then d to both targets curl -sS -X DELETE "$COUCH/a/X?rev=$REV" > /dev/null rep a d "$ONLY_X" for t in t1 t2; do rep d $t; done curl -sS "$COUCH/t1/X" curl -sS "$COUCH/t2/X" ``` Output on 3.5.3: ``` {"error":"not_found","reason":"deleted"} {"_id":"X","_rev":"1-cd6dcd82d939d5b6220025fbc573b36a","v":"a"} ``` ### Your Environment CouchDB 3.5.3 (git 7a1f97558), Erlang/OTP 28, single node, q=1 n=1, macOS arm64. ### Additional Context From reading the source, I think two things interact: - `couch_db:before_docs_update` only runs `prep_and_validate_replicated_updates` when the db has a VDU (or a design doc or attachments in the request). That step drops a replicated revision that is already a leaf; without a VDU the same path gets merged. - `couch_replicator_worker:find_missing` skips `_revs_diff` once the missing ratio average goes over 0.95 (6 batches in a row where everything was missing), and then re-sends revisions the target already has, with the longer history. The same split happens with default replication settings when 20,000 unrelated docs come before X in `a`'s changes. If the unrelated docs come after X, or there are only 5, both targets end the same. The relevant files look unchanged on `main`. Fuller repro with all runs and controls: https://github.com/blackwell-systems/normalization-confluence/tree/main/research/couchdb-replicator -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
