3. 3
Objective
• 2 fold:
• Quest for a more performant data format than
Avro for nested data
• Understand and test new data formats in general
4. 4
Hadoop data formats
• Sequence file. It stores key-value pairs of data in
a flat binary file. Rows stored as values.
• ORC. Stores column oriented data. Added RLE
and Dictionary encoding, and statistics, single file
output. Will add Bloom filter.
• Avro. Data serialization framework: serialization
format & exchange service, for any language. Data
accompanied by schema (in JSON). Supports
schema evolution.
5. 5
Parquet
• Columnar storage
• Automatic dictionary encoding and run-length
encoding. Separation of encoding vs compression.
• Run-length encoding: replaces sequences ("runs")
of consecutive repeated characters (or other units
of data) with a single character and the length of
the run.
• Dictionary encoding takes the different values
present in a column, and represents each one in
compact 2-byte form
6. 6
Parquet
• Parquet can handle multiple schemas. Support
schema evolution.
• LogType A : organizationId, userId, timestamp,
recordId, cpuTime
• LogType V : userId, organizationId, timestamp,
foo, bar
• Can be used by any project in the Hadoop
ecosystem. Integrations provided for M/R, Pig,
Hive, Cascading and Impala.
7. 7
Parquet
• SELECT vs INSERT.
• Parquet tables require relatively little memory to
query, because a query reads and decompresses data
in 8MB chunks.
• Inserting into a Parquet table is a more memory-
intensive operation because the data for each data file
(with a maximum size of 1GB) is stored in memory
until encoded, compressed, and written to disk.
8. 8
Parquet
• Memory issues (Heap space error) resolved by:
• Reducing the parquet.block.size.The block size is the
size of a row group being buffered in memory and its
default value is 256 MB.
• The total memory allocated was around 1 GB.
• Using multiple Hive partitions -> multiple buffers were
getting created (one for writing into each partition ) .
• So writing data using parquet will always have a high
memory requirement .
• Hive’s Distribute by: was workaround to memory issues!
9. 9
Parquet vs other formats
Performance test with 100G data over multiple queries
Parquet wins
10. 10
Impala overview
• MPP implementation of a query engine
• Impala vs Hive: SQL queries for interactive
exploratory analytics on large data sets. Vs Hive,
runs as batch.
• Not using M/R – but uses HDFS
• Not CEP – closer to a RDBMS.
• Impala uses the same metadata store as Hive to
record information about table structure and
properties
11. 11
Impala overview
• Can create a table in Hive, and use it in Impala
• E.g. Impala doesn’t support Avro, but Hive does
• Language is mix between SQL & HiveQL
• Requires a lot of memory (128 G min./node)
• Initial load of data via Refresh; can take a lot of time
• loads the block location data for newly added data
files
12. 12
Impala overview
• Shortcomings
• Impala doesn’t support nested types at this point
(version 1.2.3) as long as it contains only Impala-
compatible data types – it cannot contain nested types
such as array, map, or struct.
• Impala currently does not "spill to disk"
• if intermediate results being processed on a node
exceed the memory reserved for Impala on that
node.
• No Custom Serializer/Deserializer classes (SerDes)
• Impala cancels a running query if any host on which that
query is executing fails
13. 13
Impala overview
• Example. For create a PARQUET table in IMPALA there
are 3 ways:
• -> PARQUET table created in HIVE (with no nested
data types).
• -> Create and load with data a normal text table in
IMPALA:
• IMPALA> create table parquet_table_name LIKE
text_table_name STORED AS PARQUET LOCATION
/user/hdfs/..’;
• Create Parquet format table and then insert into parquet
table using normal text table.
• IMPALA> insert overwrite table parquet_table_name
select * from text_table_name;
14. 14
Use Case
• Can't query Avro table in Impala because having
nested columns.
• Avro table created through Hive, we can use it in
Impala as long as it contains only Impala-compatible
data types.
• (cannot contain nested types such as array, map,
orstruct).
15. 15
Use Case
• How to deal with nested XML data in Hadoop?
• There is no direct mapping from xml to avro. Process goes:
• Parse XML and Convert to Avro : Parse XML using XMLStreamReader and
• Perform JAXB unmarshalling and Create Avro Records from JAXB objects.Need to write
a java class for this.Tried using Parquet/Avro:
• Tested: Process Xml – first convert into Avro and then store into Parquet format using
parquet-avro apis.
• The problem is the Schema provided has some arrays which is union of type string and
null both.
• Currently this AvroSchemaConverter is not able to handle such avro schema and it gives
exception.
• Tested: Impala 1.2.3 on CDH 4.5
• Impala doesn’t support nested types at this point