Agree with Will that falling back to _all_docs-powered queries is usually undesirable in all but the smallest data sets. More folks than you'd think end up going into production without the right index because the _all_docs-powered query in development (with a small data set) seemed to be fast enough.
I always advise people to use "use_index" so they get the predictability of "this query uses that index". You're then left with the the user wondering whether index X is built yet and for that they have to navigate _active_tasks or poll a query until it returns something, which is a little primitive but probably beyond the scope of Garren's original post. On Thu, 26 Mar 2020 at 09:04, Will Holley <willhol...@gmail.com> wrote: > Broadly, I think it's a big step forward if we can prevent Mango from > automatically selecting extremely stale indexes. > > I've been going back and forth on whether step 3 could lead to some > difficult-to-predict behaviour. If we assume that requests have a short > timeout - e.g. we can't return any result if it doesn't complete within the > FDB transaction timeout - then I think it's fine: queries that use > _all_docs and a large database will be timing out anyway. > > If we were to allow long-running queries then it seems a bit sketchier > because adding an index to a large database could cause queries that > previously completed to start timing out whilst they block on the index > build. This is basically how Mango in CouchDB 2/3 behaves and has been a > big pain point for customers I've worked with, to the point where you > basically need to explicitly specify which index Mango uses in all cases if > you're to avoid surprise timeouts when somebody adds a new index. > > As I understand it, we're not allowing queries to span FDB transactions so > this latter case is not something to worry about? > > Cheers, > > Will > > On Wed, 25 Mar 2020 at 19:43, Garren Smith <gar...@apache.org> wrote: > > > On Wed, Mar 25, 2020 at 8:35 PM Paul Davis <paul.joseph.da...@gmail.com> > > wrote: > > > > > > It was therefore felt that having an immediate "Not ready" signal for > > > just _some_ calls to _find, based on the type of backing index, was a > bad > > > and confusing api. > > > > > > > > We also discussed _find calls where the user does not specify an > index, > > > and concluded that we would be free to choose between using the > _all_docs > > > index (which is always up to date but rarely the best index for a given > > > selector) or blocking to update a better but stale index. > > > > > > > > Summary-ing my summarisation; > > > > > > > > 1) if you specify an index, we'll use it even if we have to update > it, > > > no matter how long that takes. > > > > 2) if you don't specify an index, it's the dealers choice. The > details > > > here may change in point releases. > > > > > > > > > > So it seems there's still a bit of confusion on what the consensus is > > > here. The way that I had thought this would work is that we'd do > > > something like such: > > > > > > 1. If user specifies and index, use it even if we have to wait > > > 2. If an index is built that can be used, use it > > > 3. If an index is building that can be used, wait for it > > > 4. As a last resort use _all_docs > > > > > > Discussing with Garren on the PR he's of the opinion that we should > > > skip step 3 and just go directly to using _all_docs if nothing is > > > built. > > > > > > > I just want to clarify step 3. I'm ok with using an index that still > needs > > to be built as long as there is no other built index > > that can service the request. > > > > So the big thing for me is to always prefer a built index over a building > > index. In the situation where there is only 1 building index versus all > > docs I'm ok with using the building index. > > > > > > > > > > > My main assumption is that most cases where a user is creating an > > > index and then wanting to run a query with it are in the > > > design/exploration phase of learning the feature or designing an index > > > to use. In that scenario if we skip waiting it seems likely that a > > > user could easily be led to believe that an index creation "worked" > > > for their selector when in reality it was just backed by _all_docs. > > > > > > The other reason for preferring to wait for an index to finish > > > building is that the UI for the normal case of creating indexes is a > > > bit awkward. Having to run a polling loop around checking the index > > > status seems suboptimal in most cases. > > > > > > Am I missing other cases that would benefit from not waiting and just > > > using _all_docs? > > > > > > Paul > > > > > >