Обсуждение: cache type info in json_agg and friends
Currently, json_agg, jsonb_agg, json_object_agg and jsonb_object_agg do type classification on their arguments on each call to the transition function. This is quite unnecessary, as the argument types won't change. This patch remedies the defect by caching the necessary values in the aggregate state object. While this doesn't change the performance much, since these functions are essentially dominated by other bits of the processing, I think it is nevertheless worth doing. There are other areas where we might attack this, also. In particular, if one of the arguments is a record, then composite_to_json(b) will do this for every attribute of every record. However, it's much less clear to me how we can cache this information sensibly. cheers andrew
Вложения
Andrew Dunstan wrote: > Currently, json_agg, jsonb_agg, json_object_agg and jsonb_object_agg do type > classification on their arguments on each call to the transition function. > This is quite unnecessary, as the argument types won't change. This patch > remedies the defect by caching the necessary values in the aggregate state > object. Seems a reasonable idea to me. This is 9.6 only, right? What's the reason for this pattern? ! json_categorize_type(val_type,&tcategory, &outfuncoid); ! state->val_category = tcategory; ! state->val_output_func = outfuncoid; I think you could just as well call json_categorize_type() with the final pointer values, and save the two separate variables, as there is no gain in clarity or ease of reading; I mean ! json_categorize_type(tmptyp, &state->val_category, &state->val_output_func); Also, and this is not new in this patch, this code reuses a variable named "val_type" for both values and keys, which reads a bit odd. -- Álvaro Herrera http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
> Currently, json_agg, jsonb_agg, json_object_agg and jsonb_object_agg do > type classification on their arguments on each call to the transition > function. This is quite unnecessary, as the argument types won't change. > This patch remedies the defect by caching the necessary values in the > aggregate state object. +1 > > While this doesn't change the performance much, since these functions > are essentially dominated by other bits of the processing, I think it is > nevertheless worth doing. Agree After quick observation of your patch, why don't you use FmgrInfo instead of JsonAggState.val_output_func/JsonAggState.key_category? FmgrInfo could be filled by fmgr_info_cxt() in aggcontext memory context. Suppose, direct usage of FmgrInfo with FunctionCall a bit faster than OidFunctionCall. -- Teodor Sigaev E-mail: teodor@sigaev.ru WWW: http://www.sigaev.ru/
On 09/14/2015 03:42 PM, Teodor Sigaev wrote: >> Currently, json_agg, jsonb_agg, json_object_agg and jsonb_object_agg do >> type classification on their arguments on each call to the transition >> function. This is quite unnecessary, as the argument types won't change. >> This patch remedies the defect by caching the necessary values in the >> aggregate state object. > +1 > >> >> While this doesn't change the performance much, since these functions >> are essentially dominated by other bits of the processing, I think it is >> nevertheless worth doing. > Agree > > After quick observation of your patch, why don't you use FmgrInfo > instead of JsonAggState.val_output_func/JsonAggState.key_category? > FmgrInfo could be filled by fmgr_info_cxt() in aggcontext memory > context. Suppose, direct usage of FmgrInfo with FunctionCall a bit > faster than OidFunctionCall. Well, we need the category to help data_to_json(b) do its work. Nevertheless, it might be doable to pass an FmgrInfo* to datum_to_json. I'll see what I can do. cheers andrew
On 09/14/2015 04:24 PM, Andrew Dunstan wrote: > > > On 09/14/2015 03:42 PM, Teodor Sigaev wrote: >>> Currently, json_agg, jsonb_agg, json_object_agg and jsonb_object_agg do >>> type classification on their arguments on each call to the transition >>> function. This is quite unnecessary, as the argument types won't >>> change. >>> This patch remedies the defect by caching the necessary values in the >>> aggregate state object. >> +1 >> >>> >>> While this doesn't change the performance much, since these functions >>> are essentially dominated by other bits of the processing, I think >>> it is >>> nevertheless worth doing. >> Agree >> >> After quick observation of your patch, why don't you use FmgrInfo >> instead of JsonAggState.val_output_func/JsonAggState.key_category? >> FmgrInfo could be filled by fmgr_info_cxt() in aggcontext memory >> context. Suppose, direct usage of FmgrInfo with FunctionCall a bit >> faster than OidFunctionCall. > > > Well, we need the category to help data_to_json(b) do its work. > Nevertheless, it might be doable to pass an FmgrInfo* to > datum_to_json. I'll see what I can do. > > The real problem about this is that in the most important cases to improve (composite_to_json(b) and the array processing functions) we'll still end up calling fmgr_info for every attribute of every record and for every array, although not for every array element, which is what OidOutputFunctionCall does for us anyway. It's not obvious to me how to fix that, and before I put lots of effort into it I want to do some profiling to see where the time is actually being spent - I don't want to add a whole lot of code for a very marginal improvement. One thought I did have that might be worth testing is that in the case of jsonb all the micro operations might be killing us, and that it might well be faster to generate a JSON string in the aggregates and then parse that into jsonb in the final function. cheers andrew
On 09/14/2015 03:41 PM, Alvaro Herrera wrote: > Andrew Dunstan wrote: > >> Currently, json_agg, jsonb_agg, json_object_agg and jsonb_object_agg do type >> classification on their arguments on each call to the transition function. >> This is quite unnecessary, as the argument types won't change. This patch >> remedies the defect by caching the necessary values in the aggregate state >> object. > Seems a reasonable idea to me. This is 9.6 only, right? I think we can reasonably backpatch it to 9.5, which is where the jsonb functions were actually introduced. It's not at all user visible, and we're still in alpha. Seem fair? I have addressed your stylistic concerns, but I'll leave the fmgr_info question Teodor raised for another day. Before I do anything more than this I want to do some profiling to find out where the time is actually going for various workloads. cheers andrew