What failure mode is likely here? As the uploads are signed, the network payload is not corruptible from the moment its written into the HTTPS request, which places it earlier
* RAM corruption which ECC doesn't pick up. It'd be interesting to know what stats & health checks AWS run here, such as, say, low-intensity RAM checks when VM space is idle. * Any temp files buffering the blocks to HDD are being corrupted, which could happen with faulty physical disk? Is that likely? * S3 itself is in trouble. I don't see any checksum verification of disk0buffered block data before it is uploaded to S3: the files are just handed straight off to the AWS SDK. I could certainly force that through the hadoop CRC check sequence, but that complicates retransmission as well as performance. What could work would be to build the MD5 sum of each block as it is written from spark to buffer, then verify that the returned etag of that POST/PUT matches the original value, That'd to end-to-end error checking from the JVM ram all the way to S3, leaving VM ECC and S3 itself as the failure points. -Steve my old work on this: https://www.slideshare.net/steve_l/did-you-reallywantthatdata On 5 Feb 2018, at 18:41, Ryan Blue <rb...@netflix.com.INVALID<mailto:rb...@netflix.com.INVALID>> wrote: In that case, I'd recommend tracking down the node where the files were created and reporting it to EMR. On Mon, Feb 5, 2018 at 10:38 AM, Dong Jiang <dji...@dataxu.com<mailto:dji...@dataxu.com>> wrote: Thanks for the response, Ryan. We have transient EMR cluster, and we do rerun the cluster whenever the cluster failed. However, in this particular case, the cluster succeeded, not reporting any errors. I was able to null out the corrupted the column and recover the rest of the 133 columns. I do feel the issue is more than 1-2 occurrences a year. This is the second time, I am aware of the issue within a month, and we certainly don’t run as large data infrastructure compared to Netflix. I will keep an eye on this issue. Thanks, Dong From: Ryan Blue <rb...@netflix.com<mailto:rb...@netflix.com>> Reply-To: "rb...@netflix.com<mailto:rb...@netflix.com>" <rb...@netflix.com<mailto:rb...@netflix.com>> Date: Monday, February 5, 2018 at 1:34 PM To: Dong Jiang <dji...@dataxu.com<mailto:dji...@dataxu.com>> Cc: Spark Dev List <dev@spark.apache.org<mailto:dev@spark.apache.org>> Subject: Re: Corrupt parquet file We ensure the bad node is removed from our cluster and reprocess to replace the data. We only see this once or twice a year, so it isn't a significant problem. We've discussed options for adding write-side validation, but it is expensive and still unreliable if you don't trust the hardware. rb On Mon, Feb 5, 2018 at 10:28 AM, Dong Jiang <dji...@dataxu.com<mailto:dji...@dataxu.com>> wrote: Hi, Ryan, Do you have any suggestions on how we could detect and prevent this issue? This is the second time we encountered this issue. We have a wide table, with 134 columns in the file. The issue seems only impact one column, and very hard to detect. It seems you have encountered this issue before, what do you do to prevent a recurrence? Thanks, Dong From: Ryan Blue <rb...@netflix.com<mailto:rb...@netflix.com>> Reply-To: "rb...@netflix.com<mailto:rb...@netflix.com>" <rb...@netflix.com<mailto:rb...@netflix.com>> Date: Monday, February 5, 2018 at 12:46 PM To: Dong Jiang <dji...@dataxu.com<mailto:dji...@dataxu.com>> Cc: Spark Dev List <dev@spark.apache.org<mailto:dev@spark.apache.org>> Subject: Re: Corrupt parquet file If you can still access the logs, then you should be able to find where the write task ran. Maybe you can get an instance ID and open a ticket with Amazon. Otherwise, it will probably start failing the HW checks when the instance hardware is reused, so I wouldn't worry about it. The _SUCCESS file convention means that the job ran successfully, at least to the point where _SUCCESS is created. I wouldn't rely on _SUCCESS to indicate actual job success (you could do other tasks after that fail) and it carries no guarantee about the data that was written. rb On Mon, Feb 5, 2018 at 9:41 AM, Dong Jiang <dji...@dataxu.com<mailto:dji...@dataxu.com>> wrote: Hi, Ryan, Many thanks for your quick response. We ran Spark on transient EMR clusters. Nothing in the log or EMR events suggests any issues with the cluster or the nodes. We also see the _SUCCESS file on the S3. If we see the _SUCCESS file, does that suggest all data is good? How can we prevent a recurrence? Can you share your experience? Thanks, Dong From: Ryan Blue <rb...@netflix.com<mailto:rb...@netflix.com>> Reply-To: "rb...@netflix.com<mailto:rb...@netflix.com>" <rb...@netflix.com<mailto:rb...@netflix.com>> Date: Monday, February 5, 2018 at 12:38 PM To: Dong Jiang <dji...@dataxu.com<mailto:dji...@dataxu.com>> Cc: Spark Dev List <dev@spark.apache.org<mailto:dev@spark.apache.org>> Subject: Re: Corrupt parquet file Dong, We see this from time to time as well. In my experience, it is almost always caused by a bad node. You should try to find out where the file was written and remove that node as soon as possible. As far as finding out what is wrong with the file, that's a difficult task. Parquet's encoding is very dense and corruption in encoded values often looks like different data. When you see a decoding exception like this, we find it is usually that the compressed data was corrupted and is no longer valid. You can look for the page of data based on the value counter, but that's about it. Even if you could find a single record that was affected, that's not valuable because you don't know whether there is other corruption that is undetectable. There's nothing to reliably recover here. What we do in this case is find and remove the bad node, then reprocess data so we know everything is correct from the upstream source. rb On Mon, Feb 5, 2018 at 9:01 AM, Dong Jiang <dji...@dataxu.com<mailto:dji...@dataxu.com>> wrote: Hi, We are running on Spark 2.2.1, generating parquet files, like the following pseudo code df.write.parquet(...) We have recently noticed parquet file corruptions, when reading the parquet in Spark or Presto, as the following: Caused by: org.apache.parquet.io<http://org.apache.parquet.io/>.ParquetDecodingException: Can not read value at 40870 in block 0 in file file:/Users/djiang/part-00122-80f4886a-75ce-42fa-b78f-4af35426f434.c000.snappy.parquet Caused by: org.apache.parquet.io<http://org.apache.parquet.io/>.ParquetDecodingException: could not read page Page [bytes.size=1048594, valueCount=43663, uncompressedSize=1048594] in col [incoming_aliases_array, list, element, key_value, value] BINARY It appears only one column in one of the rows in the file is corrupt, the file has 111041 rows. My questions are 1) How can I identify the corrupted row? 2) What could cause the corruption? Spark issue or Parquet issue? Any help is greatly appreciated. Thanks, Dong -- Sent from: http://apache-spark-developers-list.1001551.n3.nabble.com/ --------------------------------------------------------------------- To unsubscribe e-mail: dev-unsubscr...@spark.apache.org<mailto:dev-unsubscr...@spark.apache.org> -- Ryan Blue Software Engineer Netflix -- Ryan Blue Software Engineer Netflix -- Ryan Blue Software Engineer Netflix -- Ryan Blue Software Engineer Netflix
