Uploaded image for project: 'IMPALA'
  1. IMPALA
  2. IMPALA-8394

Inconsistent data read from S3a connector

    XMLWordPrintableJSON

    Details

    • Type: Bug
    • Status: Resolved
    • Priority: Critical
    • Resolution: Not A Problem
    • Affects Version/s: Impala 3.2.0, Impala 3.3.0
    • Fix Version/s: Not Applicable
    • Component/s: Backend
    • Labels:
      None
    • Epic Color:
      ghx-label-6

      Description

      While testing a build with remote data cache (https://github.com/michaelhkw/impala/commits/remote-cache-debug) with S3, it was noticed that data read back from S3 through the HDFS S3 adaptor was inconsistent. This was confirmed by computing the checksum of the buffer right after a successful read. The following are the activities of 2 threads in the log.

      Both thread 18922 and 18924 tried to look up s3a://impala-remote-reads/tpcds_3000_decimal_parquet.db/store_sales/ss_sold_date_sk=2451097/46490ec1e79b939b-b427e7c50000000b_986349703_data.0.parq at offset: 89814317. Both of them hit cache miss. They both read from S3 for the content. Thread 18924 won the race to insert into the cache. When 18922 came around later to try to insert the same entry into the cache, it noticed that the checksum of the content inserted by thread 18924 was different from its own content.

      Please note that the checksum of the bytes read from S3 were computed and logged in hdfs-file-reader.cc before the insertion into the cache (which also computed the checksum again) and the inconsistency was also observed in hdfs-file-reader.cc already, with thread 18924 computing 8299739883147237483 while thread 18922 computing 9118051972380785265.

      We re-ran the same experiment with --use_hdfs_pread=true and the problem went away. While I don't rule out bugs in the cache prototype at this point, the debugging so far suggests the content read back from S3 via HDFS S3a connector is inconsistent when pread was disabled. It could be that we inadvertently shared the file handle somehow or there are some race conditions in the S3a connector which got exposed by the timing change with the cache enabled.

      FWIW, we also ran the same experiment in HDFS remote read configuration and it was not reproducible there either.

      Thread 18924

      I0405 12:02:15.316999 18924 data-cache.cc:344] ed4c2ab7791b5883:9f1507450000005f] Looking up s3a://impala-remote-reads/tpcds_3000_decimal_parquet.db/store_sales/ss_sold_date_sk=2451097/46490ec1e79b939b-b427e7c50000000b_986349703_data.0.parq mtime: 1549425284000 offset: 89814317 bytes_to_read: 8332914 bytes_read: 0 buffer: 4d600000
      
      I0405 12:02:15.593314 18924 hdfs-file-reader.cc:185] ed4c2ab7791b5883:9f1507450000005f] Caching file s3a://impala-remote-reads/tpcds_3000_decimal_parquet.db/store_sales/ss_sold_date_sk=2451097/46490ec1e79b939b-b427e7c50000000b_986349703_data.0.parq mtime: 1549425284000 offset: 89814317 bytes_read 8332914 checksum 8299739883147237483
      
      I0405 12:02:15.596087 18924 data-cache.cc:233] ed4c2ab7791b5883:9f1507450000005f] Storing file /data0/1/impala/datacache/cf4b57f89e5985f2:487084b5c69b208b offset 1669431296 len 8332914 checksum 8299739883147237483
      
      I0405 12:02:15.602699 18924 data-cache.cc:361] ed4c2ab7791b5883:9f1507450000005f] Storing s3a://impala-remote-reads/tpcds_3000_decimal_parquet.db/store_sales/ss_sold_date_sk=2451097/46490ec1e79b939b-b427e7c50000000b_986349703_data.0.parq mtime: 1549425284000 offset: 89814317 bytes_to_read: 8332914 buffer: 4d600000 stored: true
      

      Thread 18922:

      I0405 12:02:15.011065 18922 data-cache.cc:344] ed4c2ab7791b5883:9f150745000000da] Looking up s3a://impala-remote-reads/tpcds_3000_decimal_parquet.db/store_sales/ss_sold_date_sk=2451097/46490ec1e79b939b-b427e7c50000000b_986349703_data.0.parq mtime: 1549425284000 offset: 89814317 bytes_to_read: 8332914 bytes_read: 0 buffer: 59200000
      
      I0405 12:02:16.281126 18922 hdfs-file-reader.cc:185] ed4c2ab7791b5883:9f150745000000da] Caching file s3a://impala-remote-reads/tpcds_3000_decimal_parquet.db/store_sales/ss_sold_date_sk=2451097/46490ec1e79b939b-b427e7c50000000b_986349703_data.0.parq mtime: 1549425284000 offset: 89814317 bytes_read 8332914 checksum 9118051972380785265
      
      I0405 12:02:16.282948 18922 data-cache.cc:166] ed4c2ab7791b5883:9f150745000000da] Storing duplicated file /data0/1/impala/datacache/cf4b57f89e5985f2:487084b5c69b208b offset 1669431296 len 8332914 checksum 8299739883147237483 buffer checksum: 9118051972380785265
      
      E0405 12:02:16.282974 18922 data-cache.cc:171] ed4c2ab7791b5883:9f150745000000da] Write checksum mismatch for file /data0/1/impala/datacache/cf4b57f89e5985f2:487084b5c69b208b offset 1669431296 entry len: 8332914 store_len: 8332914 Expected 8299739883147237483, Got 9118051972380785265.
      
      I0405 12:02:16.283023 18922 data-cache.cc:361] ed4c2ab7791b5883:9f150745000000da] Storing s3a://impala-remote-reads/tpcds_3000_decimal_parquet.db/store_sales/ss_sold_date_sk=2451097/46490ec1e79b939b-b427e7c50000000b_986349703_data.0.parq mtime: 1549425284000 offset: 89814317 bytes_to_read: 8332914 buffer: 59200000 stored: false
      

      The problem is quite reproducible with TPCDS Q28 at TPCDS 3000 with parquet format.

      select  *
      from (select avg(ss_list_price) B1_LP
                  ,count(ss_list_price) B1_CNT
                  ,count(distinct ss_list_price) B1_CNTD
            from store_sales
            where ss_quantity between 0 and 5
              and (ss_list_price between 185 and 185+10 
                   or ss_coupon_amt between 10548 and 10548+1000
                   or ss_wholesale_cost between 6 and 6+20)) B1,
           (select avg(ss_list_price) B2_LP
                  ,count(ss_list_price) B2_CNT
                  ,count(distinct ss_list_price) B2_CNTD
            from store_sales
            where ss_quantity between 6 and 10
              and (ss_list_price between 28 and 28+10
                or ss_coupon_amt between 6100 and 6100+1000
                or ss_wholesale_cost between 27 and 27+20)) B2,
           (select avg(ss_list_price) B3_LP
                  ,count(ss_list_price) B3_CNT
                  ,count(distinct ss_list_price) B3_CNTD
            from store_sales
            where ss_quantity between 11 and 15
              and (ss_list_price between 173 and 173+10
                or ss_coupon_amt between 6371 and 6371+1000
                or ss_wholesale_cost between 32 and 32+20)) B3,
           (select avg(ss_list_price) B4_LP
                  ,count(ss_list_price) B4_CNT
                  ,count(distinct ss_list_price) B4_CNTD
            from store_sales
            where ss_quantity between 16 and 20
              and (ss_list_price between 101 and 101+10
                or ss_coupon_amt between 2938 and 2938+1000
                or ss_wholesale_cost between 21 and 21+20)) B4,
           (select avg(ss_list_price) B5_LP
                  ,count(ss_list_price) B5_CNT
                  ,count(distinct ss_list_price) B5_CNTD
            from store_sales
            where ss_quantity between 21 and 25
              and (ss_list_price between 8 and 8+10
                or ss_coupon_amt between 5093 and 5093+1000
                or ss_wholesale_cost between 50 and 50+20)) B5,
           (select avg(ss_list_price) B6_LP
                  ,count(ss_list_price) B6_CNT
                  ,count(distinct ss_list_price) B6_CNTD
            from store_sales
            where ss_quantity between 26 and 30
              and (ss_list_price between 110 and 110+10
                or ss_coupon_amt between 2276 and 2276+1000
                or ss_wholesale_cost between 36 and 36+20)) B6
      limit 100;
      

      cc'ing Sahil Takiar, Joe McDonnell Lars Volker Todd Lipcon David Rorke

        Attachments

          Activity

            People

            • Assignee:
              kwho Michael Ho
              Reporter:
              kwho Michael Ho
            • Votes:
              0 Vote for this issue
              Watchers:
              2 Start watching this issue

              Dates

              • Created:
                Updated:
                Resolved: