Full visit without any selection criteria is dropping handful of documents with discrepancy in documentCount

[nehajatav]

Describe the bug
Every time we run a full visit of the cluster we are losing out on few documents exactly same in number as the #documents misreported in documentCount of response. The missed documents were not consistently the same, but they were not created nor reindexed during the visit (created in 2021)

To Reproduce
Run full visit using slices
Very few visit calls results in #documents in json lesser than documentCount json key provided in visit response

Expected behavior
We expected the missed document to be there in a full visit run esp since the document was an old one and not recently created

Environment (please complete the following information):

• OS: RHEL8
• Infrastructure: Podman

Vespa version
8.332.5

[vekterli]

From my understanding this is observed when using stream=true, correct? Does this also happen if stream=false, i.e. buffered response mode?

[nehajatav]

We only use stream=true