I have create a company specific branch and added 4 new flags to control this behavior, these gave us a huge performance boost when running Spark jobs on snapshots of very large tables in S3. I tried to do everything cleanly but
a) not being familiar with the whole test strategies I haven't had time to add any useful tests, though of course I left the default behavior the same, and a lot of the behavior I control wit these flags only affect performance, not the final result, so I would need some pointers on how to add useful tests b) I added a new flag to be an overall override for prefetch behavior that overrides any setting even in the column family descriptor, not sure if what I did was entirely in the spirit of what HBase does Again these if used properly would only impact jobs using TableSnapshotInputFormat in their Spark or M-R jobs. Would someone from the core team be willing to look at my patch? I have never done this before, so would appreciate a quick pointer on how to send a patch and get some quick feedback. Cheers. ---- Saad On Sat, Mar 10, 2018 at 9:56 PM, Saad Mufti <saad.mu...@gmail.com> wrote: > The question remain though of why it is even accessing a column family's > files that should be excluded based on the Scan. And that column family > does NOT specify prefetch on open in its schema. Only the one we want to > read specifies prefetch on open, which we want to override if possible for > the Spark job. > > ---- > Saad > > > On Sat, Mar 10, 2018 at 9:51 PM, Saad Mufti <saad.mu...@gmail.com> wrote: > >> See below more I found on item 3. >> >> Cheers. >> >> ---- >> Saad >> >> On Sat, Mar 10, 2018 at 7:17 PM, Saad Mufti <saad.mu...@gmail.com> wrote: >> >>> Hi, >>> >>> I am running a Spark job (Spark 2.2.1) on an EMR cluster in AWS. There >>> is no Hbase installed on the cluster, only HBase libs linked to my Spark >>> app. We are reading the snapshot info from a HBase folder in S3 using >>> TableSnapshotInputFormat class from HBase 1.4.0 to have the Spark job read >>> snapshot info directly from the S3 based filesystem instead of going >>> through any region server. >>> >>> I have observed a few behaviors while debugging performance that are >>> concerning, some we could mitigate and other I am looking for clarity on: >>> >>> 1) the TableSnapshotInputFormatImpl code is trying to get locality >>> information for the region splits, for a snapshots with a large number of >>> files (over 350000 in our case) this causing single threaded scan of all >>> the file listings in a single thread in the driver. And it was useless >>> because there is really no useful locality information to glean since all >>> the files are in S3 and not HDFS. So I was forced to make a copy of >>> TableSnapshotInputFormatImpl.java in our code and control this with a >>> config setting I made up. That got rid of the hours long scan, so I am good >>> with this part for now. >>> >>> 2) I have set a single column family in the Scan that I set on the hbase >>> configuration via >>> >>> scan.addFamily(str.getBytes())) >>> >>> hBaseConf.set(TableInputFormat.SCAN, convertScanToString(scan)) >>> >>> >>> But when this code is executing under Spark and I observe the threads >>> and logs on Spark executors, I it is reading from S3 files for a column >>> family that was not included in the scan. This column family was >>> intentionally excluded because it is much larger than the others and so we >>> wanted to avoid the cost. >>> >>> Any advice on what I am doing wrong would be appreciated. >>> >>> 3) We also explicitly set caching of blocks to false on the scan, >>> although I see that in TableSnapshotInputFormatImpl.java it is again >>> set to false internally also. But when running the Spark job, some >>> executors were taking much longer than others, and when I observe their >>> threads, I see periodic messages about a few hundred megs of RAM used by >>> the block cache, and the thread is sitting there reading data from S3, and >>> is occasionally blocked a couple of other threads that have the >>> "hfile-prefetcher" name in them. Going back to 2) above, they seem to be >>> reading the wrong column family, but in this item I am more concerned about >>> why they appear to be prefetching blocks and caching them, when the Scan >>> object has a setting to not cache blocks at all? >>> >> >> I think I figured out item 3, the column family descriptor for the table >> in question has prefetch on open set in its schema. Now for the Spark job, >> I don't think this serves any useful purpose does it? But I can't see any >> way to override it. If these is, I'd appreciate some advice. >> > >> Thanks. >> >> >>> >>> Thanks in advance for any insights anyone can provide. >>> >>> ---- >>> Saad >>> >>> >> >> >