Hello Gopal, I have been looking further into this issue, and have found that the non-determinstic behavior of Hive in generating DAGs is actually due to the logic in AggregateStatsCache.findBestMatch() called from AggregateStatsCache.get(), as well as the disproportionate distribution of Nulls in __HIVE_DEFAULT_PARTITION__ (in the case of the TPC-DS dataset).
Here is what is happening. Let me use web_sales table and ws_web_site_sk column in the 10TB TPC-DS dataset as a running example. 1. In the course of running TPC-DS queries, Hive asks MetaStore about the column statistics of 1823 partNames in the web_sales/ws_web_site_sk combination, either without __HIVE_DEFAULT_PARTITION__ or with __HIVE_DEFAULT_PARTITION__. --- Without __HIVE_DEFAULT_PARTITION__, it reports a total of 901180 nulls. --- With __HIVE_DEFAULT_PARTITION__, however, it report a total of 1800087 nulls, almost twice as many. 2. The first call to MetaStore returns the correct result, but all subsequent requests are likely to return the same result from the cache, irrespective of the inclusion of __HIVE_DEFAULT_PARTITION__. This is because AggregateStatsCache.findBestMatch() treats __HIVE_DEFAULT_PARTITION__ in the same way as other partNames, and the difference in the size of partNames[] is just 1. The outcome depends on the duration of intervening queries, so everything is now non-deterministic. 3. If a wrong value of numNulls is returned, Hive generates a different DAG, which usually takes much longer than the correct one (e.g., 150s to 1000s for the first part of Query 24, and 40s to 120s for Query 5). I guess the problem is particularly pronounced here because of the huge number of nulls in __HIVE_DEFAULT_PARTITION__. It is ironic to see that the query optimizer is so efficient that a single wrong guess of numNulls creates a very inefficient DAG. Note that this behavior cannot be avoided by setting hive.metastore.aggregate.stats.cache.max.variance to zero because the difference in the number of partNames[] between the argument and the entry in the cache is just 1. I think that AggregateStatsCache.findBestMatch() should treat __HIVE_DEFAULT_PARTITION__ in a special way, by not returning the result in the cache if there is a difference in the inclusion of partName __HIVE_DEFAULT_PARTITION__ (or should provide the use with an option to activate this feature). However, I am testing only with the TPC-DS data, so please take my claim with a grain of salt. --- Sungwoo On Fri, Jul 20, 2018 at 2:54 PM Gopal Vijayaraghavan <gop...@apache.org> wrote: > > My conclusion is that a query can update some internal states of > HiveServer2, affecting DAG generation for subsequent queries. > > Other than the automatic reoptimization feature, there's two other > potential suspects. > > First one would be to disable the in-memory stats cache's variance param, > which might be triggering some residual effects. > > hive.metastore.aggregate.stats.cache.max.variance > > I set it to 0.0 when I suspect that feature is messing with the runtime > plans or just disable the cache entirely with > > set hive.metastore.aggregate.stats.cache.enabled=false; > > Other than that, query24 is an interesting query. > > Is probably one of the corner cases where the predicate push-down is > actually hurting the shared work optimizer. > > Also cross-check if you have accidentally loaded store_sales with > ss_item_sk(int) and if the item i_item_sk is a bigint (type mismatches will > trigger a slow join algorithm, but without any consistency issues). > > Cheers, > Gopal > > >
