phoenix-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Perko, Ralph J" <Ralph.Pe...@pnnl.gov>
Subject Re: Pig vs Bulk Load record count
Date Tue, 03 Feb 2015 23:29:55 GMT
I have solved the problem.  This was a mystery because the same data loaded into the same schema
gave conflicting counts depending on the load technique.  While the data itself had no duplicate
keys the behavior suggested something was up with the keys (MR input / output had the correct
record count for both load techniques for instance).  I confirmed this by creating a pig udf
that created a uuid for each row as the pk.  The result of running this test was each row
appeared as expected and I got the correct count.  But I couldn’t figure out why the data
itself would behave differently because it was also unique.  My pig script could hardly be
simpler with no transformations, it is a simple load and store.  This ended up being the issue!

Solution:
Assign the correct pig data type to the PK values rather than letting pig figure it out. 
I am not sure what the exact underlying issue is, but this fixed it (perhaps when pig coerced
the values to a datatype it thought best it munged it somehow).

Changes to pig script from below:

Z = load '$data' USING PigStorage(',') as (
  file_name:chararray,
  rec_num:int,

Thanks for the help
Ralph

From: <Ciureanu>, "Constantin (GfK)" <Constantin.Ciureanu@gfk.com<mailto:Constantin.Ciureanu@gfk.com>>
Reply-To: "user@phoenix.apache.org<mailto:user@phoenix.apache.org>" <user@phoenix.apache.org<mailto:user@phoenix.apache.org>>
Date: Tuesday, February 3, 2015 at 1:52 AM
To: "user@phoenix.apache.org<mailto:user@phoenix.apache.org>" <user@phoenix.apache.org<mailto:user@phoenix.apache.org>>
Subject: RE: Pig vs Bulk Load record count

Hello Ralph,

Try to check if the PIG script doesn’t produce keys that overlap (that would explain the
reduce in number of rows).

Good luck,
   Constantin

From: Ravi Kiran [mailto:maghamravikiran@gmail.com]
Sent: Tuesday, February 03, 2015 2:42 AM
To: user@phoenix.apache.org<mailto:user@phoenix.apache.org>
Subject: Re: Pig vs Bulk Load record count

Thanks Ralph. I will try to reproduce this on my end with a sample data set and get back to
you.
Regards
Ravi

On Mon, Feb 2, 2015 at 5:27 PM, Perko, Ralph J <Ralph.Perko@pnnl.gov<mailto:Ralph.Perko@pnnl.gov>>
wrote:
Ravi,

The create statement is attached.  You will see some additional fields I excluded from the
first email.

Thanks!
Ralph

________________________________
From: Ravi Kiran [maghamravikiran@gmail.com<mailto:maghamravikiran@gmail.com>]
Sent: Monday, February 02, 2015 5:03 PM
To: user@phoenix.apache.org<mailto:user@phoenix.apache.org>

Subject: Re: Pig vs Bulk Load record count

Hi Ralph,
   Is it possible to share the CREATE TABLE command as I would like to reproduce the error
on my side with a sample dataset with the specific data types of yours.
Regards
Ravi

On Mon, Feb 2, 2015 at 1:29 PM, Perko, Ralph J <Ralph.Perko@pnnl.gov<mailto:Ralph.Perko@pnnl.gov>>
wrote:
Ravi,

Thanks for the help - I am sorry I am not finding the upsert statement.  Attache are the logs
and output.  I specify the columns because I get errors if I do not.

I ran a test on 10K records.  Pig states it processed 10K records.  Select count() says 9030.
 I analyzed the 10k data in excel and there are no duplicates

Thanks!
Ralph

__________________________________________________
Ralph Perko
Pacific Northwest National Laboratory
(509) 375-2272<tel:%28509%29%20375-2272>
ralph.perko@pnnl.gov<mailto:ralph.perko@pnnl.gov>

From: Ravi Kiran <maghamravikiran@gmail.com<mailto:maghamravikiran@gmail.com>>
Reply-To: "user@phoenix.apache.org<mailto:user@phoenix.apache.org>" <user@phoenix.apache.org<mailto:user@phoenix.apache.org>>
Date: Monday, February 2, 2015 at 12:23 PM

To: "user@phoenix.apache.org<mailto:user@phoenix.apache.org>" <user@phoenix.apache.org<mailto:user@phoenix.apache.org>>
Subject: Re: Pig vs Bulk Load record count

Hi Ralph,
   Regarding the upsert query in the logs, it should be Phoenix Custom Upsert Statement: 
as you have explicitly specified the fields in STORE .    Is it possible to give it a try
with a smaller set of records , say 8k to see the behavior.
Regards
Ravi

On Mon, Feb 2, 2015 at 11:27 AM, Perko, Ralph J <Ralph.Perko@pnnl.gov<mailto:Ralph.Perko@pnnl.gov>>
wrote:
Thanks for the quick response.  Here is what I have below:

========================================
Pig script:
-------------------------------
register $phoenix_jar;

Z = load '$data' USING PigStorage(',') as (
  file_name,
  rec_num,
  epoch_time,
  timet,
  site,
  proto,
  saddr,
  daddr,
  sport,
  dport,
  mf,
  cf,
  dur,
  sdata,
  ddata,
  sbyte,
  dbyte,
  spkt,
  dpkt,
  siopt,
  diopt,
  stopt,
  dtopt,
  sflags,
  dflags,
  flags,
  sfseq,
  dfseq,
  slseq,
  dlseq,
  category);

STORE Z into 'hbase://$table_name/FILE_NAME,REC_NUM,EPOCH_TIME,TIMET,SITE,PROTO,SADDR,DADDR,SPORT,DPORT,MF,CF,DUR,SDATA,DDATA,SBYTE,DBYTE,SPKT,DPKT,SIOPT,DIOPT,STOPT,DTOPT,SFLAGS,DFLAGS,FLAGS,SFSEQ,DFSEQ,SLSEQ,DLSEQ,CATEGORY'
using org.apache.phoenix.pig.PhoenixHBaseStorage('$zookeeper','-batchSize 5000');

=========================

I cannot find the upsert statement you are referring to in either the MR logs or Pig output
but I do have this below – Pig thinks it output the correct number of records

Input(s):
Successfully read 42871627 records (1479463169 bytes) from: "/data/incoming/201501124931/SAMPLE"

Output(s):
Successfully stored 42871627 records in: "hbase://TEST/FILE_NAME,REC_NUM,EPOCH_TIME,TIMET,SITE,PROTO,SADDR,DADDR,SPORT,DPORT,MF,CF,DUR,SDATA,DDATA,SBYTE,DBYTE,SPKT,DPKT,SIOPT,DIOPT,STOPT,DTOPT,SFLAGS,DFLAGS,FLAGS,SFSEQ,DFSEQ,SLSEQ,DLSEQ,CATEGORY"


Count command:
select count(1) from TEST;

__________________________________________________
Ralph Perko
Pacific Northwest National Laboratory
(509) 375-2272<tel:%28509%29%20375-2272>
ralph.perko@pnnl.gov<mailto:ralph.perko@pnnl.gov>

From: Ravi Kiran <maghamravikiran@gmail.com<mailto:maghamravikiran@gmail.com>>
Reply-To: "user@phoenix.apache.org<mailto:user@phoenix.apache.org>" <user@phoenix.apache.org<mailto:user@phoenix.apache.org>>
Date: Monday, February 2, 2015 at 11:01 AM
To: "user@phoenix.apache.org<mailto:user@phoenix.apache.org>" <user@phoenix.apache.org<mailto:user@phoenix.apache.org>>
Subject: Re: Pig vs Bulk Load record count

Hi Ralph,
   That's definitely a cause of worry. Can you please share the UPSERT query being built by
Phoenix . You should see it in the logs with an entry "Phoenix Generic Upsert Statement: ..
Also, what do the MapReduce counters say for the job.  If possible can you share the pig script
as sometimes the order of columns in the STORE command impacts.
Regards
Ravi


On Mon, Feb 2, 2015 at 10:46 AM, Perko, Ralph J <Ralph.Perko@pnnl.gov<mailto:Ralph.Perko@pnnl.gov>>
wrote:
Hi, I’ve run into a peculiar issue between loading data using Pig vs the CsvBulkLoadTool.
 I have 42M csv records to load and I am comparing the performance.

In both cases the MR jobs are successful, and there are no errors.
In both cases the MR job counters state there are 42M Map input and output records

However, when I run count on the table when the jobs are complete something is terribly off.
After the bulk load, select count shows all 42M recs in Phoenix as is expected.
After the pig load there are only 3M recs in Phoenix – not even close.

I have no errors to send.  I have run the same test multiple times and gotten the same results.
   The pig script is not doing any transformations.  It is a simple LOAD and STORE
I get the same result using client jars from 4.2.2 and 4.2.3-SNAPSHOT.  4.2.3-SNAPSHOT is
running on the region servers.

Thanks,
Ralph






Mime
View raw message