Hi Paul and Team, I reach out to you again for suggestion on the UNION ALL performance between Parquet and database.
Observation : the query is not parallelised in multiple drillbits when we select from database. But the same number of rows are parallelised when we select from Parquet. This brings down the performance of UNION ALL between database and Parquet. I have shared few details in trailing mail. Please let me know if you need any other specific information. Appreciate your help on this. Thanks in advance, Sreeparna Bhabani On Wed, 6 May 2020, 16:42 sreeparna bhabani, <[email protected]> wrote: > Hi Paul, > > Thanks for the analysis you did. The issue occurs when we UNION ALL > between Parquet and database. Here find below more details on the scenario- > > The size of the folder- > 849 MB > The total number of parquet files under this folder - 19 > Total number of rows in this folder - 25245 K > Selected rows from Parquet after filter- 169 K > Selected rows from database - 50 > > I have tested the select query from database with same number of rows > (16402 K). But it is not parallelized for database query. In fact I tried > DB UNION ALL DB also. But it is not parallelized ( each of the DB query has > 109 K rows). > > I tried "Parquet JOIN database" for testing purpose to check whether it is > parallelized during joining. Although the JDBC_SCAN was not parallelized, > PARQUET_ROW_GROUP_SCAN parallelized here. > > FYI - The planner.slice_target = 100000 > > I have raised one JIRA ticket. Please find the same - > <https://issues.apache.org/jira/browse/DRILL-7720> > https://issues.apache.org/jira/browse/DRILL-7720 > > The query we used- > SELECT column1, column2,...column25 > FROM dfs.`root`.`Parquet` > WHERE column1 = <> AND column2 = <> AND column3 = <> > UNION ALL > SELECT column1, column2,...column25 > FROM db.`Database` > WHERE column1 = <> AND column2 = <> AND column3 = <> > > Below is the Plan after redacted names- > 00-00 Screen > 00-01 Project(Column1=[$0],...,Column25=[$24]) > 00-02 UnionExchange > 01-01 UnionAll(all=[true]) > 01-03 Project (Column1=[$0],...,Column25=[$24]) > 01-05 SelectionVectorRemover > 01-06 > Filter(condition=[AND(=$0,'<value>',),=($1,'<value>',=($2,'<value>'))]) > 01-07 Scan(table=[[]], groupscan=[ParquetGroupScan], selectionroot=<>, > numFiles=13, numRowGroups=69, usedMetadatafile=true, usedMetastore=false, > filter=booleanAnd(..), cacheFileRoot=.., columns=[.....] ) > 01-02 Project(Column1=[$0],...,Column25=[$24]) > 01-04 Jdbc() > > json- > { > "head" : { > "version":1, > "generator" : {"type":"ExplainHandler","info":""}, > "type" : "APACHE_DRILL_PHYSICAL","options" : > [],"queue":0,"hasResourcePlan":false,"resultMode":"EXEC" > }, > "graph":[ > {"pop":"jdbc-scan"}, > {"pop":"project","initialAllocation"" > 1000000,"maxAllocation":10000000000,"cost":{memoryCost}:4194304,"outputRowCount":1}, > {"pop":"parquet-scan", > "cost":{"memoryCost":4194304,"outputRowCount":1.2287038E7} }, > {"pop":"filter" ,"initialAllocation"" > 1000000,"maxAllocation":10000000000,"cost":{"memoryCost":4194304,"outputRowCount":41468.75} > }, > {"pop":"selection-vector-remover" ,"initialAllocation"" > 1000000,"maxAllocation":10000000000,"cost":{memoryCost}:4194304,"outputRowCount":41468.75 > } , > {"pop":"project" ,"initialAllocation"" > 1000000,"maxAllocation":10000000000,"cost":{"memoryCost":4194304,"outputRowCount":41468.75} > } > {"pop":"union-all" ,"initialAllocation"" > 1000000,"maxAllocation":10000000000,"cost":{"memoryCost":4194304,"outputRowCount":41468.75} > } , > {"pop":"union-exchange" ,"initialAllocation"" > 1000000,"maxAllocation":10000000000,"cost":{"memoryCost":4194304,"outputRowCount":41468.75 > } } , > {"pop":"project" ,"initialAllocation"" > 1000000,"maxAllocation":10000000000,"cost":{"memoryCost":4194304,"outputRowCount":41468.75},"outputProj":true > }, > {"pop":"screen" ,"initialAllocation"" > 1000000,"maxAllocation":10000000000,"cost":{"memoryCost":4194304,"outputRowCount":41468.75} > } > ] > } > > Please let me know if any other information you need. > > Thanks, > Sreeparna Bhabani > > On Wed, May 6, 2020 at 2:05 PM sreeparna bhabani < > [email protected]> wrote: > >> >> ---------- Forwarded message --------- >> From: Paul Rogers <[email protected]> >> Date: Tue, 5 May 2020, 13:21 >> Subject: Re: Suggestion needed for UNION ALL performance in Apache drill >> To: <[email protected]> >> Cc: Sreeparna Bhabani <[email protected]>, <[email protected]>, >> <[email protected]> >> >> >> Hi Sreeparna, >> >> >> Thanks much for digging into the details. SQL is pretty complex and >> things don't always work as we might expect. >> >> >> The first question is: which plan is correct: the (Parquet UNION ALL DB) >> plan or the (Parquet UNION ALL Parquet) plan? I tried poking around, and >> got no definitive answer on whether UNION ALL implies ordering. On the >> one hand, the parts of the standards that I could find didn't seem to imply >> ordering. A UNION (no ALL) can't imply ordering since it essentially >> does an anti-join which may be hash-partitioned. But, a StackOverflow post >> suggested that there is an implied ordering in the case of an ORDER BY on >> the sub-queries: >> >> >> (SELECT ... ORDER BY ...) >> >> UNION AL >> >> (SELECT ... ORDER BY ...) >> >> >> That is, if we can sort each sub-query, then doing so only makes sense if >> all results of the first are returned before any results of the second. >> (Have not checked if the above is valid in Drill. Even if it was, the >> planner should handle the above as a special case.) >> >> >> So, let's assume that the (Parquet UNION ALL Parquet) case is the correct >> behavior. Then, we can speculate that the planner is getting confused >> somehow in the mixed case. Each data source (file for Parquet, JDBC for the >> DB) parallelizes independently. Each decides it needs just one fragment. >> Somehow the planner must be saying, "well, if they both only want one >> fragment, let's run the whole query in a single fragment." >> >> >> Perhaps the decision is based on row count and the planner somehow thinks >> the row counts will be small for one or both of the queries. In fact, what >> happens if you do a query with (DB UNION ALL DB)? Drill's ability to >> estimate row counts is poor, especially from JDBC. Perhaps the planner is >> guessing the tables are small and parallelizing is unnecessary. >> >> >> My advice is to file a JIRA ticket with as much detail as you can >> provide. Certainly the information from your-email. Ideally, also the query >> (with names redacted if necessary.) Also, the JSON query plan (obtained by >> using EXPLAIN PLAN FOR), again with names redacted if necessary. >> >> >> With that info, we can dig a bit deeper to determine why the two cases >> come out differently. >> >> >> Thanks, >> >> - Paul >> >> >> >> On Monday, May 4, 2020, 9:39:12 AM PDT, sreeparna bhabani < >> [email protected]> wrote: >> >> >> Hi Team, >> >> After further checking on this UNION ALL, I found that UNION ALL >> (between Parquet and database) behaves as expected with limited number of >> rows and columns. But for a larger Parquet file and higher number of >> selected rows and columns, the UNION ALL takes much higher time than sum >> of >> the same of individual Parquet and DB Query. >> >> As per the analysis, it looks like the source of this issue is- >> Although we are using distributed mode, the UNION ALL query is executed >> only on 1 NODE in case of Parquet UNION ALL DB. It is not distributed and >> parallelized in multiple nodes. >> >> Whereas, for individual query or UNION ALL between same type datasets >> (Parquet + Parquet) it is getting distributed in 2 NODES. >> >> Do you have any finding / idea on this ? >> >> Thanks, >> Sreeparna Bhabani >> >> On Tue, Apr 28, 2020 at 9:00 PM sreeparna bhabani < >> [email protected]> wrote: >> >> > Hi Paul Team, >> > >> > Please check the observation mentioned in the below Jira where we found >> > that UNION ALL query is not parallelized between multiple nodes when >> there >> > are 2 types dataset (Parquet and Database). But it is parallelized if we >> > query individual Parquet file. >> > >> > Is there any way to enforce parallel execution in multiple nodes ? >> > >> > Thanks, >> > Sreeparna Bhabani >> > >> > >> > On Tue, 28 Apr 2020, 20:46 sreeparna bhabani, < >> [email protected]> >> > wrote: >> > >> >> >> >> Hi Paul and Team, >> >> >> >> As you suggested I have created a Jira ticket which is - >> >> https://issues.apache.org/jira/browse/DRILL-7720. >> >> I have mentioned details in the Jira you asked. Please have a look. As >> >> the data is sensitive, I am trying to create dummy dataset. Will >> >> provide once it is ready. >> >> >> >> Thanks, >> >> Sreeparna Bhabani >> >> >> >> On Fri, Apr 24, 2020 at 11:28 AM sreeparna bhabani < >> >> [email protected]> wrote: >> >> >> >>> >> >>> ---------- Forwarded message --------- >> >>> From: Paul Rogers <[email protected]> >> >>> Date: Thu, 23 Apr 2020, 23:59 >> >>> Subject: Re: Suggestion needed for UNION ALL performance in Apache >> drill >> >>> To: <[email protected]>, sreeparna bhabani < >> >>> [email protected]> >> >>> Cc: <[email protected]>, <[email protected]> >> >>> >> >>> >> >>> Hi Sreeparna, >> >>> >> >>> >> >>> As suggested in the earlier e-mail, we would not expect to see >> different >> >>> performance in UNION ALL than in a simple scan. Clearly you've found >> some >> >>> kind of issue. The next step is to investigate that issue, which is a >> bit >> >>> hard to do over e-mail. >> >>> >> >>> >> >>> Please file a JIRA ticket to describe the issue and provide a >> >>> reproducible test case including query and data. If your data is >> sensitive, >> >>> please create a dummy data set, or use the provided TPC-H data set to >> >>> recreate the issue. We can then take a look to see what might be >> happening. >> >>> >> >>> >> >>> Thanks, >> >>> >> >>> - Paul >> >>> >> >>> >> >>> >> >>> On Thursday, April 23, 2020, 10:18:13 AM PDT, sreeparna bhabani < >> >>> [email protected]> wrote: >> >>> >> >>> >> >>> Hi Team, >> >>> >> >>> In addition to the below mail I have another finding. Please consider >> >>> below scenarios. The first 2 scenarios are giving expected results in >> terms >> >>> of performance. But we are not getting expected performance for 3rd >> >>> scenario which is UNION ALL with 2 different types of datasets. >> >>> >> >>> *Scenario 1- Parquet UNION ALL Parquet* >> >>> Individual execution time of 1st query - 5 secs >> >>> Individual execution time of 2nd query - 5 secs >> >>> UNION ALL of both queries execution time - 10 secs >> >>> >> >>> *Scenario 2 - DB query UNION ALL DB* *query* >> >>> Individual execution time of 1st query - 5 secs >> >>> Individual execution time of 2nd query - 5 secs >> >>> UNION ALL of both queries execution time - 10 secs >> >>> >> >>> *Scenario 3 - Parquet UNION ALL DB query* >> >>> Individual execution time of 1st query - 5 secs >> >>> Individual execution time of 2nd query - 1 sec >> >>> UNION ALL execution time - 20 secs >> >>> Ideally the execution time should not be more than 6 secs. >> >>> >> >>> May I request you to check whether the UNION ALL performance of 3rd >> >>> scenario is expected with different dataset types. >> >>> >> >>> Please suggest if there is any specific way to bring down the >> execution >> >>> time of 3rd scenario. >> >>> >> >>> Thanks in advance. >> >>> >> >>> Sreeparna Bhabani >> >>> >> >>> >> >>> >> >>> On Thu, 23 Apr 2020, 12:18 sreeparna bhabani, < >> >>> [email protected]> wrote: >> >>> >> >>> Hi Team, >> >>> >> >>> Apart from the below issue I have another question. >> >>> >> >>> Is there any relation between number of row groups and performance ? >> >>> >> >>> In the below query the number of files is 13 and numRowGroups is 69. >> Is >> >>> the UNION ALL takes more time if the number of rowgroup is high like >> that. >> >>> >> >>> Please note that the individual Parquet query takes 6 secs. But UNION >> >>> ALL takes 20 secs. Details are given in trail mail. >> >>> >> >>> Thanks, >> >>> Sreeparna Bhabani >> >>> >> >>> On Thu, 23 Apr 2020, 11:08 sreeparna bhabani, <[email protected] >> > >> >>> wrote: >> >>> >> >>> Hi Paul, >> >>> >> >>> Please find the details below. We are using 2 drillbits. Heap memory >> 16 >> >>> G, Max direct memory 32 G. One query selects from Parquet. Another one >> >>> selects fron JDBC. The parquet file size is 849 MB. It is UNION ALL. >> There >> >>> is not sorting. >> >>> >> >>> Single parquet query- >> >>> Total execution time - 6.6 sec >> >>> Scan time - 0.152 sec >> >>> Screen wait time - 5.3 sec >> >>> >> >>> Single JDBC query- >> >>> Total execution time - 0.261 sec >> >>> JDBC scan - 0.152 sec >> >>> Screen wait - 0.004 sec >> >>> >> >>> >> >>> Union all query - >> >>> Execution time - 21. 118 sec >> >>> Screen wait time - 5.351 sec >> >>> Parquet scan - 15.368 sec >> >>> Unordered receiver wait time - 14.41 sec >> >>> >> >>> Thanks, >> >>> Sreeparna Bhabani >> >>> >> >>> >> >>> On Thu, 23 Apr 2020, 10:43 Paul Rogers, <[email protected]> wrote: >> >>> >> >>> Hi Sreeparna, >> >>> >> >>> >> >>> The short answer is it *should* work: a UNION ALL is simply an append. >> >>> (Be sure you are not using a plain UNION as that needs to do more >> work to >> >>> remove duplicates.) >> >>> >> >>> >> >>> Since you are seeing unexpected behavior, we may have some kind of >> issue >> >>> to investigate and perhaps fix. Always hard to do over e-mail, but >> let's >> >>> see what we can do. >> >>> >> >>> >> >>> The first question is to understand the full query: are you doing more >> >>> than a simple scan of two files and a UNION ALL? Are there sorts or >> joins >> >>> involved? >> >>> >> >>> >> >>> The best place to start to investigate performance issues is the query >> >>> profile, which it looks like you are doing. What is the time for the >> scans >> >>> if you run each of the two scans separately? You said that they take >> 8 and >> >>> 1 seconds. Is that for the whole query or just the scan operators? >> >>> >> >>> >> >>> Then, when you run the UNION ALL, again looking at the scan operators, >> >>> is there any difference in run times? If the scans take longer, that >> is one >> >>> thing to investigate. If the scans take the same amount of time, what >> other >> >>> operator(s) are taking the rest of the time? Your note suggests that >> it is >> >>> the scan taking the time. But, there should be two scan operators: >> one for >> >>> each file. How is the time divided between them? >> >>> >> >>> >> >>> How large are the data files? Using what storage system? How many >> >>> Drillbits? How much memory? >> >>> >> >>> >> >>> Thanks, >> >>> >> >>> - Paul >> >>> >> >>> >> >>> >> >>> On Wednesday, April 22, 2020, 11:32:24 AM PDT, sreeparna bhabani < >> >>> [email protected]> wrote: >> >>> >> >>> >> >>> Hi Team, >> >>> >> >>> I reach out to you for a specific problem regarding UNION ALL. There >> is >> >>> one >> >>> UNION ALL statement which combines 2 queries. The individual queries >> are >> >>> taking 8 secs and 1 sec respectively. But UNION ALL takes 30 secs. >> >>> PARQUET_SCAN_ROW_GROUP takes the maximum time. Apache drill version is >> >>> 1.17. >> >>> >> >>> Please help to suggest how to improve this UNION ALL performance. We >> are >> >>> using parquet file. >> >>> >> >>> Thanks, >> >>> Sreeparna Bhabani >> >>> >> >>> >> >> >> >> -- >> >> >> >> Thanks n Regards, >> >> >> *Sreeparna Bhabani* >> >> >> > >> >> -- >> >> Thanks n Regards, >> *Sreeparna Bhabani* >> > > > -- > > Thanks n Regards, > *Sreeparna Bhabani* >
