_all_docs performance degrades as doc_del_count increases
---------------------------------------------------------
Key: COUCHDB-1076
URL: https://issues.apache.org/jira/browse/COUCHDB-1076
Project: CouchDB
Issue Type: Bug
Reporter: Nathan Vander Wilt
The time required to query _all_docs?limit=1 can be proportional to the
doc_del_count of the database depending on where the deleted docs are in the
sort order. As kocolosk explained on IRC:
The tree that serves as the source for _all_docs contains a small record for
each document, deleted or not ... if you have a large number of deleted docs
and their IDs are interspersed with the non-deleted ones i can imagine that it
would cause additional seeks when streaming the _all_docs response
In my use case (https://github.com/natevw/RQMS) and in other cases (e.g. a
rolling log or any long-lived database), the deleted docs may all be at the
beginning of the _all_docs view, making query performance end up like using
"?limit=N&skip=doc_del_count".
To improve the performance in cases where large blocks of documents have been
deleted, kocolosk notes:
[10:30am] kocolosk: the inner nodes in the btree currently report the doc_count
and doc_del_count
[10:31am] kocolosk: we might be able to rewrite the function that walks the
btree so that it checks if the doc_count underneath an inner node is zero
[10:31am] kocolosk: and then it can skip that part of the tree entirely
[10:31am] kocolosk: instead of descending all the way to the leaf nodes and
skipping deleted documents one by one
[10:31am] n[ate]vw: yeah, that'd have the benefit of being a code-only change
[10:31am] kocolosk: right
[10:32am] kocolosk: i think it should work rather nicely, though it probably
requires some interesting spelunking into the depths of couch_btree
[10:33am] kocolosk: davisp: whaddya think? would what i'm proposing be
possible? (checking for doc_count=0 in the inner node reduction and then
skipping ahead when serving _all_docs)
...
[10:39am] davisp: n[ate]vw: Its definitely doable, its just a question of how
to do it so that I don't bleed from the eyes when trying to read through the
implementation
...
[10:40am] rnewson: davisp: right. we definitely want to preserve the high
readability of the current btree code. /s
[10:40am] davisp: kocolosk: The only thing I can think of is passing a function
to the view iteration that gets called to evaluate whether it should decend to
a child node based on the key/reduction pair
[10:41am] kocolosk: davisp: what you're proposing would also allow for (more)
efficient implementation of skip
--
This message is automatically generated by JIRA.
-
For more information on JIRA, see: http://www.atlassian.com/software/jira
