We have a JIRA for this issue that hopefully will be fixed in the next release. https://issues.apache.org/jira/browse/DRILL-5089
Thanks, Padma On Feb 17, 2017, at 1:50 PM, David Kincaid <[email protected]<mailto:[email protected]>> wrote: My apologies for not following up sooner. Earlier this week our DevOps engineer was looking into this problem as well and discovered the root cause of our issue. We developed a custom storage provider that utilizes S3 as the pstore. We thought this was just storing configuration information (esp. storage plugin config), but we discovered that it was spending a lot of time reading files in a /temp/drill subdirectory of our S3 bucket. We removed the custom plugin and things are running much better now. I have one of our engineers working on this now to see where we went wrong. My question for the list now is if you know what exactly it is doing. We really want to be able to store the storage plugin config on S3 so that it is persisted between restarts of the EMR cluster that we are running Drill on. If you have any suggestions or advice it would be much appreciate. I really appreciate all the time and patience you all showed helping us troubleshoot this issue. I'm glad in the end that it really was something on our end and not something more mysterious happening in Drill itself. Thanks, Dave On Wed, Feb 15, 2017 at 12:37 PM, Jinfeng Ni <[email protected]<mailto:[email protected]>> wrote: Can you help try one more thing if you can? Run jstack on the foreman Drillbit process while the query is doing the query planning. Capture the jstack every one second or couple of seconds consecutively for some time, by appending the jstack output into one log file. Take a look at the stack trace for the forman thread in the form of "275b623b-bb15-8bd8-fd29-f9a571a7534e:foreman" (The first part is the query ID). If foreman thread is stuck in one method call, it may show up in the log repeatedly. In this way we may have a better idea what's the cause of the problem. Based on the tests you tried, the combination of the query / parquet files probably hit a bug in the code that we are not aware of currently. Without the parquet files to re-reproduce, it's hard to debug the issue and find a possible fix. On Wed, Feb 15, 2017 at 8:35 AM, David Kincaid <[email protected]<mailto:[email protected]>> wrote: I ran that EXPLAIN that you suggested against the entire 100 file table and it takes about 3 seconds. I will try to get a defect written up in the next few days. - Dave On Tue, Feb 14, 2017 at 9:06 PM, Jinfeng Ni <[email protected]<mailto:[email protected]>> wrote: >From the two tests you did, I'm inclined to think there might be some special things in your parquet files. How do you generate these parquet files? Do they contain normal data type (int/float/varchar), or complex type (array/map)? In our environment, we also have hundreds of parquet files, each with size ~ hundreds of MBs. A typical query (several tables joined) would takes a couple of seconds in planning. One more test if you can help run. EXPLAIN PLAN FOR SELECT someCol1, someCol2 FROM dfs.`parquet/transaction/OneSingleFile.parquet`; The above query is simple enough that planner should not spend long time in enumerating different choices. If it still takes long time for query planning, the more likely cause might be in parquet files you used. On Tue, Feb 14, 2017 at 1:06 PM, David Kincaid <[email protected]<mailto:[email protected]>> wrote: I will write up a defect. The first test you suggested below - running the query on just one of our Parquet files produces the same result (10-12 minute planning time). However, the second test - using cp.`tpch/nation.parquet` - results in a planning time of only about a minute. So, I'm not sure how to interpret that. What does that mean to you all? - Dave On Tue, Feb 14, 2017 at 12:37 PM, Jinfeng Ni <[email protected]<mailto:[email protected]>> wrote: Normally, the slow query planning could be caused by : 1. Some planner rule hit a bug when processing certain operators in the query, for instance join operator, distinct aggregate. The query I tried on a small file seems to rule out this possibility. 2. The parquet metadata access time. According to the long, this does not seem to be the issue. 3. Something we are not aware of. To help get some clue, can you help do the following: 1. run the query over one single parquet files, in stead of 100 parquet files? You can change using dfs.`parquet/transaction/OneSingleFile.parquet`. I'm wondering if the planning time is proportional to # of parquet files. 2. What if you try your query by replacing dfs.`parquet/transaction/OneSingleFile.parquet` with cp.`tpch/nation.parquet` which is a small tpch parquet file (you need re-enable the storage plugin 'cp')? Run EXPLAIN should be fine. This will tell us if the problem is caused by the parquet source, or the query itself. Yes, please create a defect in Drill JIRA. On Tue, Feb 14, 2017 at 5:02 AM, David Kincaid < [email protected]<mailto:[email protected]>> wrote: Thank you for the feedback. It seems there is nothing more I can do on my end. What are my next steps? Shall I create a defect in the Drill Jira? - Dave On Mon, Feb 13, 2017 at 5:13 PM, Jinfeng Ni <[email protected]<mailto:[email protected]>> wrote: The size of parquet files will matter in terms of meta data access time, which is just 212 ms according to your log file. My understanding is it does not matter too much to the overall planning times. That's why it probably makes sense to try over such a small toy example. Normally the planning time for such simple query should be much shorter than 12 minutes. It indicates it could be caused by a code bug, or something else that we are currently unaware of. On Mon, Feb 13, 2017 at 2:47 PM, David Kincaid < [email protected]<mailto:[email protected]>> wrote: The example in DRILL-5183 is just a very small toy example to demonstrate the bug with how Drill reads Parquet array fields. It doesn't have anything to do with this planning issue (at least I don't think it does). Sorry if I confused things with that reference. I just tried running our query directly against the table at dfs.`parquet/transaction` and get the same result (12 minutes of planning time). I disabled the cp and s3 storage plugins that were enabled so that only the dfs storage plugin is enabled and the result is the same. Is this expected for Drill to take this long in the planning phase for a query? Is there anything else I can try or information I could provide to help identify the bug (seems like a bug to me)? I really appreciate you guys helping out so quickly this afternoon. - Dave On Mon, Feb 13, 2017 at 4:13 PM, Jinfeng Ni <[email protected]<mailto:[email protected]>> wrote: I downloaded books.parquet from DRILL-5183, and created a view on top of this single parquet file. Then, run EXPLAIN for the query, and it completes within 1.2 seconds on Drill 1.8.0 release. (The # of parquet files would impact the time to fetch metadata. Since it's not the bottleneck in this case, it should not cause a big difference). Do you see the long planning time issue for this query only, or it happens for other queries as well? Besides the possibility of planning rule bugs, we once saw another possible cause of long planning issue. In your storage plugin configuration, if you enable some other storage plugin (for instance, hbase, or hive etc) which are slow to access, then those un-relevant storage plugin might impact your query as well. You may temporarily disable those storage plugins, and see if it's the cause of the problem. 0: jdbc:drill:zk=local> explain plan for . . . . . . . . . . . > select fltb1.sapId, yearmo, . . . . . . . . . . . > COUNT(*) as totalcnt, . . . . . . . . . . . > count(distinct(CASE . . . . . . . . . . . > WHEN . . . . . . . . . . . > (REPEATED_CONTAINS(fltb1. classLabels, . . . . . . . . . . . > 'Thing:Service:MedicalService:Diagnostic:Radiology: Ultrasound.*')) . . . . . . . . . . . > THEN fltb1.invoiceId . . . . . . . . . . . > END)) as ultracount, . . . . . . . . . . . > count(distinct (CASE . . . . . . . . . . . > WHEN . . . . . . . . . . . > (REPEATED_CONTAINS(fltb1. classLabels, . . . . . . . . . . . > 'Thing:Service:MedicalService:Diagnostic:LaboratoryTest.*')) . . . . . . . . . . . > THEN fltb1.invoiceId . . . . . . . . . . . > END)) as labcount . . . . . . . . . . . > from ( . . . . . . . . . . . > select sapid, invoiceId, . . . . . . . . . . . > TO_CHAR(TO_TIMESTAMP( transactionDate, 'YYYY-MM-dd HH:mm:ss.SSSSSS'), 'yyyy-MM') yearmo, . . . . . . . . . . . > classLabels . . . . . . . . . . . > from dfs.tmp.transactionView) fltb1 . . . . . . . . . . . > group by fltb1.sapId, yearmo; +------+------+ | text | json | +------+------+ | 00-00 Screen 00-01 Project(sapId=[$0], yearmo=[$1], totalcnt=[$2], ultracount=[$3], labcount=[$4]) .................................... 00-09 SelectionVectorRemover 00-12 Sort(sort0=[$0], sort1=[$1], dir0=[ASC], dir1=[ASC]) 00-15 HashAgg(group=[{0, 1}], totalcnt=[COUNT()]) ................................ 00-22 Scan(groupscan=[ ParquetGroupScan [entries=[ReadEntryWithPath [path=file:/tmp/parquet/ transaction]], selectionRoot=file:/tmp/parquet/transaction, numFiles=1, usedMetadataFile=false, columns=[`sapId`, `invoiceId`, `transactionDate`, `classLabels`.`array`]]]) 1 row selected (1.195 seconds) On Mon, Feb 13, 2017 at 1:51 PM, David Kincaid < [email protected]<mailto:[email protected]>> wrote: Here is the entire transactionView.view.drill file. As you can see the view itself is very simple and is just wrapping a syntactic problem with the array field. That's an issue I reported in Jira under DRILL-5183 ( https://issues.apache.org/jira/browse/DRILL-5183) { "name" : "transactionView", "sql" : "SELECT `transactionRowKey`, `sapId`, `practiceName`, `practiceCity`, `practiceState`, `practicePostalCode`, `animalId`, `dateOfBirth`, `species`, `breed`, `gender`, `status`, `ownerId`, `itemType`, `classification`, `subclass`, `practiceDescription`, `clientDescription`, `invoiceId`, `unitOfMeasure`, `vendorName`, `vaccine`, `rabies`, `vaccineType`, `price`, `quantity`, `transactionDate`, `visitReason`, `speciesCode`, `genderCode`, `t`.`classLabels`['array'] AS `classLabels`\nFROM `dfs`.`/parquet/transaction` AS `t`", "fields" : [ { "name" : "transactionRowKey", "type" : "ANY", "isNullable" : true }, { "name" : "sapId", "type" : "ANY", "isNullable" : true }, { "name" : "practiceName", "type" : "ANY", "isNullable" : true }, { "name" : "practiceCity", "type" : "ANY", "isNullable" : true }, { "name" : "practiceState", "type" : "ANY", "isNullable" : true }, { "name" : "practicePostalCode", "type" : "ANY", "isNullable" : true }, { "name" : "animalId", "type" : "ANY", "isNullable" : true }, { "name" : "dateOfBirth", "type" : "ANY", "isNullable" : true }, { "name" : "species", "type" : "ANY", "isNullable" : true }, { "name" : "breed", "type" : "ANY", "isNullable" : true }, { "name" : "gender", "type" : "ANY", "isNullable" : true }, { "name" : "status", "type" : "ANY", "isNullable" : true }, { "name" : "ownerId", "type" : "ANY", "isNullable" : true }, { "name" : "itemType", "type" : "ANY", "isNullable" : true }, { "name" : "classification", "type" : "ANY", "isNullable" : true }, { "name" : "subclass", "type" : "ANY", "isNullable" : true }, { "name" : "practiceDescription", "type" : "ANY", "isNullable" : true }, { "name" : "clientDescription", "type" : "ANY", "isNullable" : true }, { "name" : "invoiceId", "type" : "ANY", "isNullable" : true }, { "name" : "unitOfMeasure", "type" : "ANY", "isNullable" : true }, { "name" : "vendorName", "type" : "ANY", "isNullable" : true }, { "name" : "vaccine", "type" : "ANY", "isNullable" : true }, { "name" : "rabies", "type" : "ANY", "isNullable" : true }, { "name" : "vaccineType", "type" : "ANY", "isNullable" : true }, { "name" : "price", "type" : "ANY", "isNullable" : true }, { "name" : "quantity", "type" : "ANY", "isNullable" : true }, { "name" : "transactionDate", "type" : "ANY", "isNullable" : true }, { "name" : "visitReason", "type" : "ANY", "isNullable" : true }, { "name" : "speciesCode", "type" : "ANY", "isNullable" : true }, { "name" : "genderCode", "type" : "ANY", "isNullable" : true }, { "name" : "classLabels", "type" : "ANY", "isNullable" : true } ], "workspaceSchemaPath" : [ ] } On Mon, Feb 13, 2017 at 3:47 PM, Jinfeng Ni <[email protected]<mailto:[email protected]>> wrote: Yes, the log confirmed that the planning, especially physical planning, is the one that took most of the time. If the definition of view s3.cisexport.transactionView is not very complicated (involves large # of tables), then it's possible that some planner rules have a bug. (In the past, we once saw couple of planner rules would be fired in a loop). Is it possible that you can share the DDL of the view? That may help us re-produce the problem and take a look at the trace of Calcite, which Drill uses as the query planner.
