Efficient	
  processing	
  of	
  large	
  and	
  
complex	
  XML	
  documents	
  in	
  Hadoop	
  
	
  
Sujoe	
  Bose	
  
Senior	
  Principal,	
  
Sabre	
  Holdings	
  
June,	
  2013	
  
Presenta.on	
  Outline	
  
§  MoBvaBon	
  
§  ETL	
  vs.	
  ELT	
  
§  Avro	
  Format	
  
§  Mapping	
  from	
  XML	
  to	
  Avro	
  
§  Interfaces	
  to	
  access	
  Avro	
  
§  Performance	
  and	
  Storage	
  consideraBons	
  
§  Other	
  types	
  of	
  storage/processing	
  formats	
  
confidenBal	
   2	
  
You	
  will	
  learn	
  about	
  …	
  
§  A	
  method	
  to	
  store	
  and	
  process	
  complex	
  XML	
  data	
  in	
  
Hadoop	
  as	
  Avro	
  files	
  
§  Interfaces	
  to	
  access	
  and	
  analyze	
  data	
  in	
  Avro	
  from	
  
Hive,	
  Java	
  and	
  Pig	
  
§  VariaBons	
  of	
  the	
  method	
  and	
  their	
  relaBve	
  trade-­‐offs	
  
in	
  storage	
  and	
  processing	
  
confidenBal	
   3	
  
Mo.va.on	
  
§  Prevalence	
  of	
  XML	
  and	
  its	
  derivaBves	
  
–  Spurred	
  by	
  WebServices	
  and	
  SOA	
  
–  Preferred	
  communicaBon	
  format	
  unBl	
  newer	
  formats	
  
entered	
  
–  Data	
  and	
  logs	
  represented	
  in	
  XML	
  
§  XML	
  –	
  metadata	
  combined	
  data	
  	
  
–  Flexibility	
  vs.	
  Complexity	
  
§  Could	
  be	
  arbitrarily	
  nested	
  and	
  large	
  
§  Volumes	
  of	
  documents	
  –	
  Big	
  Data	
  
confidenBal	
   4	
  
Challenges	
  
§  Parsing	
  XML	
  is	
  CPU	
  Intensive	
  
§  Certain	
  parsers/parsing	
  methods	
  result	
  in	
  more	
  
memory	
  consumpBon	
  
§  Repeated	
  parsing	
  for	
  each	
  query	
  
§  Large	
  and	
  deeply	
  nested	
  XMLs	
  makes	
  problem	
  worse	
  
§  Presence	
  of	
  tags	
  in	
  data	
  result	
  in	
  high	
  I/O	
  due	
  to	
  
storage	
  size	
  
§  Special	
  handling	
  of	
  opBonal	
  fields	
  
confidenBal	
   5	
  
ETL	
  vs.	
  ELT	
  
confidenBal	
   6	
  
§  Hadoop	
  generally	
  built	
  for	
  EL	
  –	
  T	
  
–  aka	
  Schema-­‐on-­‐Read	
  
–  Load	
  as-­‐is	
  
–  Transform	
  on	
  Access/Query	
  
§  Compare	
  with	
  Data	
  Warehouse	
  ETL	
  
–  Aka	
  Schema-­‐on-­‐Write	
  
–  Transform	
  and	
  Load	
  
–  Queries	
  are	
  lot	
  simpler	
  
–  TransformaBon	
  and	
  cleansing	
  done	
  a	
  priori	
  
Mix	
  of	
  ETL	
  and	
  ELT	
  
§  Generally	
  beaer	
  in	
  
Flexibility	
  
§  More	
  suitable	
  for	
  simpler	
  
and	
  well-­‐defined	
  formats	
  
§  More	
  applicable	
  for	
  
experimentaBon	
  
§  XML	
  data	
  parsed	
  on	
  
demand	
  for	
  every	
  query	
  
confidenBal	
   7	
  
§  Generally	
  beaer	
  in	
  
Performance	
  
§  More	
  suitable	
  when	
  
substanBal	
  cleansing	
  and	
  
reformacng	
  is	
  needed	
  
§  RepeBBve	
  queries	
  and	
  
producBon	
  workloads	
  
§  XML	
  Data	
  pre-­‐parsed	
  to	
  
minimize	
  resource	
  usage	
  
ELT	
   ETL	
  
Approaches	
  
confidenBal	
   8	
  
XML	
  Files	
  
Avro	
  Files	
  
ETL	
  
Pre-­‐parsing	
  
Pig	
  
UDF	
  
Avro	
  
Schema	
  
On-­‐demand	
  
Parsing	
  
Interfaces	
  Processing	
  Data	
  
Hive	
  
SerDe	
  
MapReduce	
   Pig	
  
UDF	
  
Hive	
  
SerDe	
  
MapReduce	
  
ELT	
  
confidenBal	
   9	
  confidenBal	
   9	
  
XML	
  Files	
  
Avro	
  Files	
  
ETL	
  
Pre-­‐parsing	
  
Pig	
  
UDF	
  
Avro	
  
Schema	
  
On-­‐demand	
  
Parsing	
  
Interfaces	
  Processing	
  Data	
  
Hive	
  
SerDe	
  
MapReduce	
   Pig	
  
UDF	
  
Hive	
  
SerDe	
  
MapReduce	
  
ETL	
  
confidenBal	
   10	
  confidenBal	
   10	
  
XML	
  Files	
  
Avro	
  Files	
  
ETL	
  
Pre-­‐parsing	
  
Pig	
  
UDF	
  
Avro	
  
Schema	
  
On-­‐demand	
  
Parsing	
  
Interfaces	
  Processing	
  Data	
  
Hive	
  
SerDe	
  
MapReduce	
   Pig	
  
UDF	
  
Hive	
  
SerDe	
  
MapReduce	
  
XML	
  Pre-­‐parsing	
  
§  Nested	
  Elements	
  and	
  Aaributes	
  
§  RepresentaBon	
  of	
  parsed	
  XML	
  Structure	
  
§  Enter	
  Avro!	
  
confidenBal	
   11	
  
Avro	
  
§  Data	
  serializaBon	
  system	
  
§  Specifically	
  designed	
  for	
  Hadoop,	
  but	
  used	
  in	
  other	
  
environments	
  also	
  
§  Rich	
  data	
  structures:	
  Arrays,	
  Records,	
  Maps	
  etc.	
  
§  Compact,	
  fast,	
  binary	
  data	
  format	
  
§  Metadata	
  stored	
  at	
  file	
  level	
  –	
  not	
  record	
  level	
  
§  	
  Split-­‐able	
  –	
  Ideal	
  for	
  Map-­‐Reduce	
  
confidenBal	
   12	
  
Avro	
  APIs	
  
§  Generic	
  Objects	
  and	
  Pre-­‐generated	
  Objects	
  
–  Easy	
  API	
  including	
  simple	
  gets	
  and	
  puts	
  
§  APIs	
  in	
  several	
  languages	
  
–  Java	
  
–  C#	
  
–  C/C++	
  
–  Python	
  
–  Ruby	
  
confidenBal	
   13	
  
Use-­‐case	
  
§  FIXML	
  –	
  Financial	
  InformaBon	
  eXchange	
  
–  hap://www.fixprotocol.org/specificaBons/	
  
§  XML	
  Database	
  Benchmark	
  
–  hap://tpox.sourceforge.net/	
  
§  Provides	
  sample	
  data	
  for	
  benchmarking	
  
§  Data	
  Generator	
  for	
  generaBng	
  large	
  and	
  predictable	
  
datasets	
  
confidenBal	
   14	
  
FIXML	
  
§  XML	
  Data	
  Generator	
  
–  hap://tpox.sourceforge.net/tpoxdata.htm	
  
§  Order:	
  Buy	
  and	
  sell	
  order	
  of	
  securiBes	
  
confidenBal	
   15	
  
Simple	
  mapping	
  
confidenBal	
   16	
  
XML	
   Avro	
   Pig	
  
Elements	
  with	
  repeated	
  
nested	
  elements	
  
Array	
   Bag	
  
Elements	
  with	
  aaributes	
  and	
  
text	
  elements	
  
Record	
   Tuple	
  
Aaributes	
  and	
  Text	
  Elements	
   Field	
   Field	
  
Avro	
  Schema	
  
{
"type": "record",
"name": "FIXOrder",
"namespace": "com.sabre.fixml",
"doc": "Definition and mapping for FIX Orders",
"mapping": "/FIXML",
"fields":
[
{ "name":"v", "type":"string", "mapping":"@v"},
{ "name":"r", "type":"string", "mapping":"@r"},
{ "name":"s", "type":"string", "mapping":"@s"},
{ "name":"Order", "mapping":"Order", "type":
{
"name":"OrderRecord", "mapping":"Order", "type": "record", "fields":
[
{ "name":"ID", "type":"string", "mapping":"@ID"},
{ "name":"ID2", "type":"string", "mapping":"@ID2"},
{ "name":"OrignDt", "type":"string", "mapping":"@OrignDt"},
{ "name":"TrdDt", "type":"string", "mapping":"@TrdDt"},
{ "name":"Acct", "type":"string", "mapping":"@Acct"},
{ "name":"AcctTyp", "type":"string", "mapping":"@AcctTyp"},
{ "name":"DayBkngInst", "type":"string", "mapping":"@DayBkngInst"},
{ "name":"BkngUnit", "type":"string", "mapping":"@BkngUnit"},
{ "name":"PreallocMeth", "type":"string", "mapping":"@PreallocMeth"},
{ "name":"AllocID", "type":"string", "mapping":"@AllocID"},
{ "name":"CshMgn", "type":"string", "mapping":"@CshMgn"},
{ "name":"ClrFeeInd", "type":"string", "mapping":"@ClrFeeInd"},
...
	
  
confidenBal	
   17	
  
Pig	
  Schema	
  
FIXOrder: tuple (
v: chararray,
r: chararray,
s: chararray,
Order: tuple (
ID: chararray,
ID2: chararray,
OrignDt: chararray,
TrdDt: chararray,
Acct: chararray,
AcctTyp: chararray,
DayBkngInst: chararray,
BkngUnit: chararray,
PreallocMeth: chararray,
AllocID: chararray,
CshMgn: chararray,
ClrFeeInd: chararray,
confidenBal	
   18	
  
Avro	
  –	
  Access	
  Methods	
  
§  Direct	
  support	
  for	
  access	
  from	
  Hive	
  (using	
  SerDe)	
  
	
  
CREATE EXTERNAL TABLE <TableName>!
ROW FORMAT SERDE
‘org.apache.hadoop.hive.serde2.avro.AvroSerDe’!
STORED as INPUTFORMAT
‘org.apache.hadoop.hive.ql.io.avro.AvroContainerInputFormat’!
OUTPUTFORMAT!
‘org.apache.hadoop.hive.ql.io.avro.AvroContainerOutputFormat’!
LOCATION ‘location-of-avro-files’!
TBLPROPERTIES ('avro.schema.url'=‘location-of-schema-
file.avsc')	
  
§  Access	
  via	
  Pig	
  -­‐	
  AvroStorage	
  
§  Avro	
  API	
  -­‐	
  Java	
  MapReduce	
  
confidenBal	
   19	
  
Test	
  Data	
  
§  Base	
  SecuriBes	
  Order	
  file	
  500,000	
  records	
  
§  Replicated	
  for	
  volume	
  
–  15x	
  -­‐	
  7.5	
  million	
  records	
  
–  30x	
  -­‐	
  15	
  million	
  records	
  
–  45x	
  -­‐	
  22.5	
  million	
  records	
  
–  60x	
  –	
  30	
  million	
  records	
  
–  75x	
  –	
  37.5	
  million	
  records	
  
	
  
confidenBal	
   20	
  
Comparison	
  
confidenBal	
   21	
  
XML	
  Files	
  
Avro	
  Files	
  
ETL	
  
Pre-­‐parsing	
  
Pig	
  
UDF	
  
Avro	
  
Schema	
  
On-­‐demand	
  
Parsing	
  
Interfaces	
  Processing	
  Data	
  
Hive	
  
SerDe	
  
MapReduce	
   Pig	
  
UDF	
  
Hive	
  
SerDe	
  
MapReduce	
  
File	
  sizes:	
  Orders	
  
§  Base	
  Data	
  
–  XML	
  file	
  size	
  as	
  is:	
  749,337,916	
  (750MB)	
  	
  
–  Gzip	
  Compressed:	
  182,687,654	
  (183MB)	
  	
  
§  Applied	
  Avro	
  conversion	
  
–  Avro	
  Snappy:	
  151,647,926	
  (152MB)	
  	
  
–  Avro	
  Gzip:	
  107,898,177	
  (108MB)	
  	
  
confidenBal	
   22	
  
Storage	
  Size	
  Comparison	
  
confidenBal	
   23	
  
Test	
  Environment	
  
§  18	
  Nodes	
  
§  Node	
  configuraBon:	
  
–  12	
  cores	
  per	
  node	
  
–  48GB	
  memory	
  
–  	
  36	
  TB	
  with	
  12	
  disks	
  of	
  3TB	
  each	
  
§  CDH	
  4.1.2	
  
confidenBal	
   24	
  
Sample	
  Query	
  
§  Security	
  Orders	
  per	
  Account	
  
order_records	
  =	
  LOAD	
  '$AVRO_INPUT'	
  using	
  AVRO_LOAD	
  AS	
  (	
  
-­‐-­‐-­‐-­‐-­‐-­‐-­‐	
  Pig	
  Schema	
  goes	
  here	
  -­‐-­‐-­‐-­‐-­‐-­‐-­‐-­‐-­‐-­‐	
  
);	
  
	
  
order_projecBon	
  =	
  FOREACH	
  order_records	
  GENERATE	
  Order.Acct	
  as	
  Account,	
  Order.OrdQty.Qty	
  
as	
  QuanBty;	
  
	
  
order_group	
  =	
  GROUP	
  order_projecBon	
  BY	
  Account;	
  
	
  
order_count	
  =	
  FOREACH	
  order_group	
  GENERATE	
  group,	
  SUM(order_projecBon.QuanBty);	
  
	
  
STORE	
  order_count	
  INTO	
  '$PIG_OUTPUT'	
  Using	
  PigStorage(',');	
  
confidenBal	
   25	
  
Run	
  Types	
  
§  Pre-­‐parsed	
  approach:	
  
–  XML	
  to	
  Avro	
  materializaBon:	
  xml-­‐to-­‐avro	
  
•  XML	
  to	
  Avro	
  is	
  run	
  only	
  once	
  on	
  the	
  data	
  
–  Avro	
  to	
  Pig	
  via	
  UDF:	
  avro-­‐to-­‐pig	
  
§  Parse	
  on	
  demand	
  
–  XML	
  parsing	
  using	
  Pig	
  UDF:	
  xml-­‐to-­‐pig	
  
confidenBal	
   26	
  
confidenBal	
   27	
  
Run	
  .me	
  in	
  Seconds	
  
Analysis	
  on	
  raw	
  XML:	
  
XML	
  to	
  Pig	
  
Pre-­‐parsing	
  XML:	
  
XML	
  to	
  Avro	
  
Analysis	
  on	
  parsed	
  XML:	
  
Avro	
  to	
  Pig	
  
confidenBal	
   28	
  
CPU	
  Usage	
  Comparison	
  
Analysis	
  on	
  raw	
  XML:	
  
XML	
  to	
  Pig	
  
Pre-­‐parsing	
  XML:	
  
XML	
  to	
  Avro	
  
Analysis	
  on	
  parsed	
  XML:	
  
Avro	
  to	
  Pig	
  
confidenBal	
   29	
  confidenBal	
   29	
  
Memory	
  Usage	
  Comparison:	
  Total	
  Memused	
  (GB)	
  
Analysis	
  on	
  raw	
  XML:	
  
XML	
  to	
  Pig	
  
Pre-­‐parsing	
  XML:	
  
XML	
  to	
  Avro	
  
Analysis	
  on	
  parsed	
  XML:	
  
Avro	
  to	
  Pig	
  
Results	
  
§  Analysis	
  on	
  pre-­‐parsed	
  data	
  compared	
  raw	
  XML	
  
–  RunBme	
  reducBon	
  by	
  more	
  than	
  50%	
  
–  Memory	
  and	
  CPU	
  consumpBon	
  reduced	
  by	
  about	
  50%	
  
§  Pre-­‐parsing	
  stage	
  takes	
  more	
  resources	
  and	
  Bme	
  
than	
  on-­‐demand	
  parsing	
  
§  RepeBBve	
  queries	
  will	
  benefit	
  from	
  one-­‐Bme	
  pre-­‐
parsing	
  
confidenBal	
   30	
  
Caveats	
  
§  Not	
  all	
  fields	
  were	
  extracted	
  from	
  the	
  XML	
  input	
  
(opBonal	
  elements)	
  
§  Challenge	
  in	
  keeping-­‐up	
  with	
  versions/changes	
  of	
  
XML	
  
§  Performance	
  numbers	
  can	
  depend	
  on	
  the	
  type	
  of	
  
data	
  and	
  the	
  mapping	
  used	
  
confidenBal	
   31	
  
Alterna.ves	
  
§  Formats	
  other	
  than	
  Avro	
  may	
  be	
  more	
  suitable	
  
§  Record	
  Columnar	
  formats	
  (RC	
  Files	
  &	
  ORC	
  Files)	
  
§  Trevni:	
  a	
  column	
  file	
  format	
  supporBng	
  Avro	
  
§  Parquet:	
  another	
  columnar	
  storage	
  for	
  Hadoop	
  
confidenBal	
   32	
  
Mo.va.on	
  for	
  Columnar	
  Format	
  
§  Map	
  Reduce	
  capability	
  
§  Column	
  ProjecBons	
  reduce	
  I/O	
  
§  Column	
  Compression	
  due	
  to	
  similarity	
  of	
  data	
  
further	
  reduces	
  I/O	
  
confidenBal	
   33	
  
Summary	
  
§  Materialized	
  version	
  well-­‐suited	
  for	
  repeated	
  queries	
  
§  For	
  ad-­‐hoc/experimental	
  queries	
  parse-­‐on-­‐demand	
  
is	
  beaer	
  
§  Mapping	
  from	
  XML	
  to	
  Avro	
  can	
  be	
  automated	
  
§  Hive,	
  Pig	
  and	
  MapReduce	
  Interfaces	
  to	
  access	
  Avro	
  
Files	
  
§  RelaBve	
  trade-­‐offs	
  between	
  flexibility	
  and	
  
performance/storage	
  
confidenBal	
   34	
  
Ques.ons	
  &	
  Comments	
  
confidenBal	
   35	
  
Thanks	
  for	
  Listening	
  
	
  sujoe.bose@sabre.com	
  
	
  

Efficient processing of large and complex XML documents in Hadoop

  • 1.
    Efficient  processing  of  large  and   complex  XML  documents  in  Hadoop     Sujoe  Bose   Senior  Principal,   Sabre  Holdings   June,  2013  
  • 2.
    Presenta.on  Outline   § MoBvaBon   §  ETL  vs.  ELT   §  Avro  Format   §  Mapping  from  XML  to  Avro   §  Interfaces  to  access  Avro   §  Performance  and  Storage  consideraBons   §  Other  types  of  storage/processing  formats   confidenBal   2  
  • 3.
    You  will  learn  about  …   §  A  method  to  store  and  process  complex  XML  data  in   Hadoop  as  Avro  files   §  Interfaces  to  access  and  analyze  data  in  Avro  from   Hive,  Java  and  Pig   §  VariaBons  of  the  method  and  their  relaBve  trade-­‐offs   in  storage  and  processing   confidenBal   3  
  • 4.
    Mo.va.on   §  Prevalence  of  XML  and  its  derivaBves   –  Spurred  by  WebServices  and  SOA   –  Preferred  communicaBon  format  unBl  newer  formats   entered   –  Data  and  logs  represented  in  XML   §  XML  –  metadata  combined  data     –  Flexibility  vs.  Complexity   §  Could  be  arbitrarily  nested  and  large   §  Volumes  of  documents  –  Big  Data   confidenBal   4  
  • 5.
    Challenges   §  Parsing  XML  is  CPU  Intensive   §  Certain  parsers/parsing  methods  result  in  more   memory  consumpBon   §  Repeated  parsing  for  each  query   §  Large  and  deeply  nested  XMLs  makes  problem  worse   §  Presence  of  tags  in  data  result  in  high  I/O  due  to   storage  size   §  Special  handling  of  opBonal  fields   confidenBal   5  
  • 6.
    ETL  vs.  ELT   confidenBal   6   §  Hadoop  generally  built  for  EL  –  T   –  aka  Schema-­‐on-­‐Read   –  Load  as-­‐is   –  Transform  on  Access/Query   §  Compare  with  Data  Warehouse  ETL   –  Aka  Schema-­‐on-­‐Write   –  Transform  and  Load   –  Queries  are  lot  simpler   –  TransformaBon  and  cleansing  done  a  priori  
  • 7.
    Mix  of  ETL  and  ELT   §  Generally  beaer  in   Flexibility   §  More  suitable  for  simpler   and  well-­‐defined  formats   §  More  applicable  for   experimentaBon   §  XML  data  parsed  on   demand  for  every  query   confidenBal   7   §  Generally  beaer  in   Performance   §  More  suitable  when   substanBal  cleansing  and   reformacng  is  needed   §  RepeBBve  queries  and   producBon  workloads   §  XML  Data  pre-­‐parsed  to   minimize  resource  usage   ELT   ETL  
  • 8.
    Approaches   confidenBal  8   XML  Files   Avro  Files   ETL   Pre-­‐parsing   Pig   UDF   Avro   Schema   On-­‐demand   Parsing   Interfaces  Processing  Data   Hive   SerDe   MapReduce   Pig   UDF   Hive   SerDe   MapReduce  
  • 9.
    ELT   confidenBal  9  confidenBal   9   XML  Files   Avro  Files   ETL   Pre-­‐parsing   Pig   UDF   Avro   Schema   On-­‐demand   Parsing   Interfaces  Processing  Data   Hive   SerDe   MapReduce   Pig   UDF   Hive   SerDe   MapReduce  
  • 10.
    ETL   confidenBal  10  confidenBal   10   XML  Files   Avro  Files   ETL   Pre-­‐parsing   Pig   UDF   Avro   Schema   On-­‐demand   Parsing   Interfaces  Processing  Data   Hive   SerDe   MapReduce   Pig   UDF   Hive   SerDe   MapReduce  
  • 11.
    XML  Pre-­‐parsing   § Nested  Elements  and  Aaributes   §  RepresentaBon  of  parsed  XML  Structure   §  Enter  Avro!   confidenBal   11  
  • 12.
    Avro   §  Data  serializaBon  system   §  Specifically  designed  for  Hadoop,  but  used  in  other   environments  also   §  Rich  data  structures:  Arrays,  Records,  Maps  etc.   §  Compact,  fast,  binary  data  format   §  Metadata  stored  at  file  level  –  not  record  level   §   Split-­‐able  –  Ideal  for  Map-­‐Reduce   confidenBal   12  
  • 13.
    Avro  APIs   § Generic  Objects  and  Pre-­‐generated  Objects   –  Easy  API  including  simple  gets  and  puts   §  APIs  in  several  languages   –  Java   –  C#   –  C/C++   –  Python   –  Ruby   confidenBal   13  
  • 14.
    Use-­‐case   §  FIXML  –  Financial  InformaBon  eXchange   –  hap://www.fixprotocol.org/specificaBons/   §  XML  Database  Benchmark   –  hap://tpox.sourceforge.net/   §  Provides  sample  data  for  benchmarking   §  Data  Generator  for  generaBng  large  and  predictable   datasets   confidenBal   14  
  • 15.
    FIXML   §  XML  Data  Generator   –  hap://tpox.sourceforge.net/tpoxdata.htm   §  Order:  Buy  and  sell  order  of  securiBes   confidenBal   15  
  • 16.
    Simple  mapping   confidenBal   16   XML   Avro   Pig   Elements  with  repeated   nested  elements   Array   Bag   Elements  with  aaributes  and   text  elements   Record   Tuple   Aaributes  and  Text  Elements   Field   Field  
  • 17.
    Avro  Schema   { "type":"record", "name": "FIXOrder", "namespace": "com.sabre.fixml", "doc": "Definition and mapping for FIX Orders", "mapping": "/FIXML", "fields": [ { "name":"v", "type":"string", "mapping":"@v"}, { "name":"r", "type":"string", "mapping":"@r"}, { "name":"s", "type":"string", "mapping":"@s"}, { "name":"Order", "mapping":"Order", "type": { "name":"OrderRecord", "mapping":"Order", "type": "record", "fields": [ { "name":"ID", "type":"string", "mapping":"@ID"}, { "name":"ID2", "type":"string", "mapping":"@ID2"}, { "name":"OrignDt", "type":"string", "mapping":"@OrignDt"}, { "name":"TrdDt", "type":"string", "mapping":"@TrdDt"}, { "name":"Acct", "type":"string", "mapping":"@Acct"}, { "name":"AcctTyp", "type":"string", "mapping":"@AcctTyp"}, { "name":"DayBkngInst", "type":"string", "mapping":"@DayBkngInst"}, { "name":"BkngUnit", "type":"string", "mapping":"@BkngUnit"}, { "name":"PreallocMeth", "type":"string", "mapping":"@PreallocMeth"}, { "name":"AllocID", "type":"string", "mapping":"@AllocID"}, { "name":"CshMgn", "type":"string", "mapping":"@CshMgn"}, { "name":"ClrFeeInd", "type":"string", "mapping":"@ClrFeeInd"}, ...   confidenBal   17  
  • 18.
    Pig  Schema   FIXOrder:tuple ( v: chararray, r: chararray, s: chararray, Order: tuple ( ID: chararray, ID2: chararray, OrignDt: chararray, TrdDt: chararray, Acct: chararray, AcctTyp: chararray, DayBkngInst: chararray, BkngUnit: chararray, PreallocMeth: chararray, AllocID: chararray, CshMgn: chararray, ClrFeeInd: chararray, confidenBal   18  
  • 19.
    Avro  –  Access  Methods   §  Direct  support  for  access  from  Hive  (using  SerDe)     CREATE EXTERNAL TABLE <TableName>! ROW FORMAT SERDE ‘org.apache.hadoop.hive.serde2.avro.AvroSerDe’! STORED as INPUTFORMAT ‘org.apache.hadoop.hive.ql.io.avro.AvroContainerInputFormat’! OUTPUTFORMAT! ‘org.apache.hadoop.hive.ql.io.avro.AvroContainerOutputFormat’! LOCATION ‘location-of-avro-files’! TBLPROPERTIES ('avro.schema.url'=‘location-of-schema- file.avsc')   §  Access  via  Pig  -­‐  AvroStorage   §  Avro  API  -­‐  Java  MapReduce   confidenBal   19  
  • 20.
    Test  Data   § Base  SecuriBes  Order  file  500,000  records   §  Replicated  for  volume   –  15x  -­‐  7.5  million  records   –  30x  -­‐  15  million  records   –  45x  -­‐  22.5  million  records   –  60x  –  30  million  records   –  75x  –  37.5  million  records     confidenBal   20  
  • 21.
    Comparison   confidenBal  21   XML  Files   Avro  Files   ETL   Pre-­‐parsing   Pig   UDF   Avro   Schema   On-­‐demand   Parsing   Interfaces  Processing  Data   Hive   SerDe   MapReduce   Pig   UDF   Hive   SerDe   MapReduce  
  • 22.
    File  sizes:  Orders   §  Base  Data   –  XML  file  size  as  is:  749,337,916  (750MB)     –  Gzip  Compressed:  182,687,654  (183MB)     §  Applied  Avro  conversion   –  Avro  Snappy:  151,647,926  (152MB)     –  Avro  Gzip:  107,898,177  (108MB)     confidenBal   22  
  • 23.
    Storage  Size  Comparison   confidenBal   23  
  • 24.
    Test  Environment   § 18  Nodes   §  Node  configuraBon:   –  12  cores  per  node   –  48GB  memory   –   36  TB  with  12  disks  of  3TB  each   §  CDH  4.1.2   confidenBal   24  
  • 25.
    Sample  Query   § Security  Orders  per  Account   order_records  =  LOAD  '$AVRO_INPUT'  using  AVRO_LOAD  AS  (   -­‐-­‐-­‐-­‐-­‐-­‐-­‐  Pig  Schema  goes  here  -­‐-­‐-­‐-­‐-­‐-­‐-­‐-­‐-­‐-­‐   );     order_projecBon  =  FOREACH  order_records  GENERATE  Order.Acct  as  Account,  Order.OrdQty.Qty   as  QuanBty;     order_group  =  GROUP  order_projecBon  BY  Account;     order_count  =  FOREACH  order_group  GENERATE  group,  SUM(order_projecBon.QuanBty);     STORE  order_count  INTO  '$PIG_OUTPUT'  Using  PigStorage(',');   confidenBal   25  
  • 26.
    Run  Types   § Pre-­‐parsed  approach:   –  XML  to  Avro  materializaBon:  xml-­‐to-­‐avro   •  XML  to  Avro  is  run  only  once  on  the  data   –  Avro  to  Pig  via  UDF:  avro-­‐to-­‐pig   §  Parse  on  demand   –  XML  parsing  using  Pig  UDF:  xml-­‐to-­‐pig   confidenBal   26  
  • 27.
    confidenBal   27   Run  .me  in  Seconds   Analysis  on  raw  XML:   XML  to  Pig   Pre-­‐parsing  XML:   XML  to  Avro   Analysis  on  parsed  XML:   Avro  to  Pig  
  • 28.
    confidenBal   28   CPU  Usage  Comparison   Analysis  on  raw  XML:   XML  to  Pig   Pre-­‐parsing  XML:   XML  to  Avro   Analysis  on  parsed  XML:   Avro  to  Pig  
  • 29.
    confidenBal   29  confidenBal   29   Memory  Usage  Comparison:  Total  Memused  (GB)   Analysis  on  raw  XML:   XML  to  Pig   Pre-­‐parsing  XML:   XML  to  Avro   Analysis  on  parsed  XML:   Avro  to  Pig  
  • 30.
    Results   §  Analysis  on  pre-­‐parsed  data  compared  raw  XML   –  RunBme  reducBon  by  more  than  50%   –  Memory  and  CPU  consumpBon  reduced  by  about  50%   §  Pre-­‐parsing  stage  takes  more  resources  and  Bme   than  on-­‐demand  parsing   §  RepeBBve  queries  will  benefit  from  one-­‐Bme  pre-­‐ parsing   confidenBal   30  
  • 31.
    Caveats   §  Not  all  fields  were  extracted  from  the  XML  input   (opBonal  elements)   §  Challenge  in  keeping-­‐up  with  versions/changes  of   XML   §  Performance  numbers  can  depend  on  the  type  of   data  and  the  mapping  used   confidenBal   31  
  • 32.
    Alterna.ves   §  Formats  other  than  Avro  may  be  more  suitable   §  Record  Columnar  formats  (RC  Files  &  ORC  Files)   §  Trevni:  a  column  file  format  supporBng  Avro   §  Parquet:  another  columnar  storage  for  Hadoop   confidenBal   32  
  • 33.
    Mo.va.on  for  Columnar  Format   §  Map  Reduce  capability   §  Column  ProjecBons  reduce  I/O   §  Column  Compression  due  to  similarity  of  data   further  reduces  I/O   confidenBal   33  
  • 34.
    Summary   §  Materialized  version  well-­‐suited  for  repeated  queries   §  For  ad-­‐hoc/experimental  queries  parse-­‐on-­‐demand   is  beaer   §  Mapping  from  XML  to  Avro  can  be  automated   §  Hive,  Pig  and  MapReduce  Interfaces  to  access  Avro   Files   §  RelaBve  trade-­‐offs  between  flexibility  and   performance/storage   confidenBal   34  
  • 35.
    Ques.ons  &  Comments   confidenBal   35   Thanks  for  Listening    sujoe.bose@sabre.com