> > ...what if we required chunk_size to be explicitly specified when the > queryset has prefetch lookups?
This seems reasonable, but would you do in the case chunk_size isn't explicitly defined - throw an exception? Currently it silently fails to prefetch which means N + 1 queries, so even prefetching for the default chunk_size of 2000 would be a huge gain in cases where chunk_size isn't defined. On Sun, 13 Jan 2019 at 02:05, charettes <charett...@gmail.com> wrote: > Replying to my concerns about the P * C queries. > > Throwing that idea out there but what if we required chunk_size to be > explicitly specified when the queryset has prefetch lookups? > > That would at least force the developers to consider how many prefetching > queries iterating through the results would require. Plus since the > argument was only recently introduced it is less likely to silently cause > changes under developers feet. > > Simon > > Le vendredi 26 octobre 2018 20:12:02 UTC-4, charettes a écrit : >> >> Josh, I agree that silently not working is problematic but it has been >> this way since prefetch_related() was introduced. >> >> Something to keep in mind as well is that silently turning it on would >> also perform P * C extra queries where P is the number of prefetches >> requested through prefetch_related() and C the number of chunks the results >> contains. This is non negligible IMO. >> >> What I'd be in favor off is raising an exception on >> prefetch_related(*prefetches).iterator() in the next release at least to >> allow users using this pattern to adjust their code and then ship the >> optimization with all the warnings related to the interactions between >> prefetch_related(*prefetches) and iterator(chunk_size) in the next one. >> >> I'm not completely completely against skipping the exception release >> phase entirely given there might be users out there accessing attributes >> expected to be prefetched on iterated instances and inadvertently >> performing tons of queries but the exception phase just feels safer given >> iterator() is usually used in memory constrained contexts. >> >> Simon >> >> Le vendredi 26 octobre 2018 19:27:55 UTC-4, Josh Smeaton a écrit : >>> >>> I tend to agree with Tobi. Prefetching silently not working on iterator >>> can be quite confusing, unless you have a good understanding of both APIs. >>> It might be possible to do what you're asking, but it'd mean that django is >>> now actually caching the result when it explicitly says it isn't - even if >>> the result is a much smaller moving cache. Prefetching chunk_size results >>> per chunk is unlikely to make a material difference to memory usage. Users >>> are usually concerned about the entire result set of the primary table. >>> >>> I don't know if you can change the API to make these suggested changes >>> without also impacting how we iterate over result sets - but I'd be >>> interested in seeing a proof of concept at the very least. >>> >>> >>> >>> On Monday, 15 October 2018 20:41:13 UTC+11, tobias....@truffls.com >>> wrote: >>>> >>>> Thank you for your feedback. I would like to answer some statements to >>>> either convince you or make it more clear, where my idea stems from: >>>> >>>> The fundamental reason why iterator() cannot be used with >>>>> prefetch_related() is that the latter requires a set of model instance to >>>>> be materialized to work appropriately which chunk_size doesn't control at >>>>> all. >>>>> In other words chunk_size only controls how many rows should be >>>>> fetched from the database cursor and kept into memory at a time. Even when >>>>> this parameter is used, iterator() will only materialize a single model >>>>> instance per yield. >>>>> >>>> >>>> It should be easily possible to change the involved code of >>>> ModelIterable to materialize the retrieved rows in batches. After >>>> materializing the batch / chunk, it could do the prefetching. >>>> >>>> >>>>> Given that iterator() always ignored prefetch related lookups instead >>>>> of erroring out when they were specified make me feel like turning such a >>>>> feature on by default could be problematic as it could balloon the memory >>>>> usage which is the main reason why iterator is useful anyway. >>>>> >>>> >>>> I would argue, that users who thoughtlessly applied prefetching >>>> together with iterator now actually get, what they thought of: less DB >>>> query round trips traded against a little more memory usage. >>>> >>>> Best, >>>> Tobi >>>> >>> -- > You received this message because you are subscribed to the Google Groups > "Django developers (Contributions to Django itself)" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to django-developers+unsubscr...@googlegroups.com. > To post to this group, send email to django-developers@googlegroups.com. > Visit this group at https://groups.google.com/group/django-developers. > To view this discussion on the web visit > https://groups.google.com/d/msgid/django-developers/1f5d8b5a-8c7c-4575-b904-adfaa9d1900d%40googlegroups.com > <https://groups.google.com/d/msgid/django-developers/1f5d8b5a-8c7c-4575-b904-adfaa9d1900d%40googlegroups.com?utm_medium=email&utm_source=footer> > . > For more options, visit https://groups.google.com/d/optout. > -- Adam -- You received this message because you are subscribed to the Google Groups "Django developers (Contributions to Django itself)" group. To unsubscribe from this group and stop receiving emails from it, send an email to django-developers+unsubscr...@googlegroups.com. To post to this group, send email to django-developers@googlegroups.com. Visit this group at https://groups.google.com/group/django-developers. To view this discussion on the web visit https://groups.google.com/d/msgid/django-developers/CAMyDDM2t87iTC3A%3DJy3qD%2Bs_5r3%3DMKNN40B2SQzcwMW6Q5DAGA%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
