CHANGES.txt - cascading - Cascading is a feature rich API for defining and executing complex and fault tolerant data processing workflows on a Hadoop cluster.

Source path: svn/ tags/ cascading-0.10.0/ CHANGES.txt

‹r579

r609

Cascading Change Log

0.10.0

Changed c.p.CoGroup "repeat" parameter to numSelfJoins to respresent the actual number of self joins to be performed.
Thus a value of 1, will cause a single self join of a pipe. Users will need to decrement the current value by 1.

Fixed bug with temporary filename generation where path created was too long.

Fixed Janino c.o.expression operations to require parameter names and types. Janino
was returning guessed parameter names in an undeterministic order.

Fixed boolean type c.t.Tuple serialization.

Fixed c.p.GroupBy merging case where grouping field names were not properly resolved.

Changed c.o.r.RegexParser to emit variable sized Tuples if a fieldDeclaration is not given. Also will emit group
matches if they are any, otherwise the match is emitted.

Removed deprecated classes; c.o.t.Texts, c.o.r.Regexes, c.p.EndPipe.

Removed experimental c.p.EndPipe class.

Changed c.t.Tap#isUseTapCollector to Tap#isWriteDirect.

Changed c.t.Tap and c.f.Flow to return c.t.TupleEntryIterator instead of c.t.TupleIterator. This is more consistent
and more useful.

Added c.t.TemplateTap to support dynamically writing out c.t.Tuple values to unique directories.

Changed Cascading to support null values returned from c.t.Tap#source() and subsequently c.t.Scheme#source().
This allows for Schemes to skip records returned by an internal Hadoop InputFormat without having to implement
a custom Hadoop InputFormat or instrument a pipe assembly with a c.o.Filter.

0.9.0

Updated c.o.Debug to allow for printing field names and tuple values in intervals.

Changed planner to fail if traps are not contained within single Map or Reduce tasks. This prevents the chance of
multiple tasks writing to the same output location. Hadoop only partially supports appends, so it is not currently
possible to append subsequent jobs to existing trap files. Naming sections of a pipe assembly allows traps to be
bound to smaller sections of assemblies.

c.o.f.Sample and c.o.f.Limit Filters. Sample allows a given percentage of Tuples to pass. Limit only allows the
specified number of Tuples to pass.

c.p.Pipe instances now capture line numbers and classnames where they are instantiated so this information
can be printed out during planner failures.

Added c.f.FlowSkipStrategy interface to allow for pluggable rules for when to skip executing a c.f.Flow participating
in a c.c.Cascade. The default implementation is c.f.FlowSkipIfSinkStale, with an optional c.f.FlowSkipIfSinkExists.
Setting a skip strategy on a Cascade overrides all Flow instance strategies.

Updated c.o.Debug to allow for printing field names and tuple values in intervals.

c.o.f.Sample and c.o.f.Limit Filters. Sample allows a given percentage of Tuples to pass. Limit only allows the
specified number of Tuples to pass.

c.p.Pipe instances now capture line numbers and classnames where they are instantiated so this information
can be printed out during planner failures.

Fixed bug with c.t.Tuple#remove() method not correctly removing values from Tuple.

Updated c.t.Tap api to support c.t.SinkMode enums. This opens up ability to support appends in the near future.

Added support for Hadoop 0.19.x. This release skips Hadoop 0.18.x.

Changed project structure so that XML functions live in their own sub-project. This includes renaming the base
Cascading tree and jars to 'core'.

Fixed bug that prevented Fields.UNKNOWN input sources from begin fed into a c.p.CoGroup for joining.

Changed all operations so that incoming c.t.Tuple and c.t.TupleEntry instances are unmodifiable. An
UnsupportedOperationException will be thrown on any attempt to modify argument tuples within an operation.
This enforces the rule argument tuples should not be modified to protect against concurrent modification in
parallel threads.

Updated c.o.r.RegexMatcher base class to use j.u.r.Matcher#find() instead of #matches(). This is more consistent
with default behaviors of popular languages. Matcher is now also initialized in prepare() and reset() in
the operation to reduce overhead.

Added new lifecycle methods to c.o.Operation, prepare and cleanup. These methods are called so that an Operation
instance can initialize and destroy any resources. They may be called more than once before the instance is
garbage collected.

Added a new operation called c.o.Buffer. Buffers are similiar to Reduce in MapReduce. They are given an Iterator
of input arguments and can emit any number of result c.t.Tuple instances. For many problems, this is more
efficient than using an c.o.Aggregator operation. Only one c.p.Every pipe with a Buffer operation may
follow a GroupBy or CoGroup.

Fixed dot file writing so GraphViz can properly load.

Upgraded jgrapht library, requires JDK 1.6.

Fixed bug where selecting postions from a c.t.Fields.UNKNOWN declaration would return the first position, not
the specified position.

Renamed c.t.Fields.KEYS to c.t.Fields.GROUP to be consistent with the Cascading model.

Fixed bug where c.t.Tap may inappropriately delete a sink from a task.

Changed c.o.Aggregator to no longer use a Map for the context. Users can now specify custom types by returning
either a new instance from start() or recycling an instance passed into start(). This change will break all existing
implementations of Aggregator. Note, simply setting a new Map<Object,Object> on the call instance in start()
should be sufficient.

Changed all c.o.Function, c.o.Filter, c.o.Aggregator, c.o.ValueAssertion, and c.o.GroupAssertions to accept
a c.f.FlowProcess object on all relevant methods. FlowProcess provides call-backs into the underyling system
to get configuration properties, fire a "keep alive" ping, or increment a custom counter. This change will
break all existing implemenations of the above interfaces.

Added ability to set serialization tokens via the cascading.serialization.tokens property. This compliments the
c.t.h.SerializationToken annotation.

Optimized co-grouping operation by using c.t.IndexTuple instead of a nested c.t.Tuple.

Changed c.t.Tap and c.s.Scheme sink methods to take a c.t.TupleEntry, instead of c.t.Fields and c.t.Tuple
individually.

Added the c.t.h.SerializationToken Java Annotation. This allows for an int value to be written during serialization
instead of a Class name for custom objects nested in c.t.Tuple instances. This feature should dramatically reduce
the size of Tuples saved in SequenceFiles, and improve the general performance during 'shuffling' between Map and
Reduce stages.

Added c.t.h.TupleSerialization, a Hadoop Serialization implementation. Tuple is no longer Hadoop Writable
and now relies on TupleSerialization for serialization support. Subequently nested objects in c.t.Tuple
only need to be c.l.Comparable. So they can be serialized properly, a Serialization implementation must be
registered with Hadoop. Note all primitive types are handled directly by Tuple, but custom types must
have a Serialization implementation, or must be Hadoop WritableComparable so that the default WritableSerialization
implementation will write them out.

0.8.3

Fix for c.p.CoGroup declared fields being generated out of order.

0.8.2

Added new properties via c.f.FlowConnector.setJarClass and c.f.FlowConnector.setJarPath for
setting the application jar file.

Fixed bug where job jar was not being inherited by subsequent MapReduce jobs when the first job was executed
in local mode.

Fixed bug where unserializable Operations were being squashed internally. c.f.Flow instances will now
fail immediately and be marked as 'failed'.

0.8.1

Fixed bug where c.t.Lfs did not force local mode for current MapReduce step.

Fixed bug where writing to a c.t.TupleCollector would fail if using a c.s.SequenceFile in some cases.

Added a few minor improvements to reduce stray object creations, and speedup c.t.Tuple serialization.

0.8.0

Updated c.o.x.TagSoupParser to accept 'features', use these features to recover past behaviors.

Updated janino and tagsoup libraries to 2.5.15 and 1.2, respectively. Note that tagsoup, in theory, is not
backwards compatible by default. See their release notes: http://home.ccil.org/~cowan/XML/tagsoup/#1.2

Added some forward compatible changes for supporting Hadoop 0.18 at the API level. Currently there are other
issues preventing some tests from passing on Hadoop 0.18.

Changed c.f.FlowException to return the parent c.f.Flow name.

Changed behavior of c.f.MultiMapReducePlanner to use c.t.h.MultiInputFormat to allow single Mappers
to support many different Hadoop InputFormat types simultaneously. This deprecates the need to normalize
sources to a map and reduces the number of jobs in a c.f.Flow in some cases.

Changed behavior of Cascading to allow for multiple paths from the same c.t.Tap source to be co-grouped on
via c.p.CoGroup. This allows for a kind of self-join where each stream is processed by a different operation
path within the Mapper.

Added c.o.f.And, c.o.f.Or, c.o.f.Xor, and c.o.f.Not logic operator c.o.Filter implementations. They should be used
to compose more complex filters from existing implementations.

Changed the behavior of c.o.BaseOperation to properly initialize itself if it is a c.o.Filter instance. This
removes the requirement that Filter implementations must set declaredFields to Fields.ALL, as it makes no
sense for a Filter to declare fields.

Added c.f.PlannerException, a subclass of c.f.FlowException, and updated c.f.MultiMapReducePlanner to throw
it on failures. Functionality of writing DOT files has been moved from FlowException to PlannerException.

Added c.o.f.FilterNotNull and c.o.f.FilterNull filter classes.

Changed c.f.MultiMapReducePlanner to fail if it encounters an c.p.Each to c.p.Every chain. In these cases, a
c.p.Group type must be between them.

Deleted c.o.Cut class as it was effectively a duplicate of c.o.Identity.

Changed c.f.MultiMapReducePlanner to fail if a c.p.GroupAssertion is not accompanied by another c.o.Aggregator
operation. This is required so that the GroupAssertion does not change the passing tuple stream if it is planned out.

Changed c.f.MultiMapReducePlanner to no longer insert new c.p.Each( ..., new Identity(), ... ) as a place holder.

Renamed c.p.PipeAssembly to c.p.SubAssembly to better reflect its purpose, which is to encapuslate reusable
pipe assemblies in the same manner as a sub-process or sub-routine. A temporary c.p.PipeAssembly class has been
provided for backwards compatibility.

Fixed bug where c.t.TapCollector would throw an NPE if a custom Tap was not using paths.

Changed behavior of c.f.Flow where if a c.f.FlowListener throws an exception, the Flow instance receiving the
exception will stop (by calling Flow.stop()). Listeners will continue to fire as expected and Flow.complete()
will re-throw the thrown exception (as was the original behavior).

Added ability to set a Cascading specific temporary directory path for use by intermediate taps created
within c.f.Flow instances. Use c.t.Hfs.setTemporaryDirectory() to configure.

Fixed bug where the 'mapred.jar' property was begin stepped on if previously set by the calling application.

Changed c.t.Tap and c.f.Flow to return c.t.TupleIterator and c.t.TupleCollector instead of c.t.TapIterator and
c.t.TapCollector, respectively.

Added c.t.Tap.flowInit( c.f.Flow flow ) to allow a given tap to know what flows it is participating in. It is called
immediately after the Flow instance is initailized.

Fixed bug with nested c.p.PipeAssembly instances where some nested assemblies threw an internal error from
the planner.

Changed c.o.Debug to accept a prefix text string that will be prefixed to every message.

Fixed bug where c.f.MultiMapReducePlanner would fail when normalizing inputs to a group where the inputs
passed through one or more splits.

Fixed bug where c.g.CoGroup silently stepped on input pipes with the same input name.

0.7.1

Fixed bug in c.f.MultiMapReducePlanner where a source used on more than one c.p.Group would cause an internal
error during planning.

Changed c.f.MultiMapReducePlanner to normalize heterogeneous sinks.

Changed c.f.MultiMapReducePlanner to keep a splitting c.p.Each on the previous step, instead of being duplicated
on each branch. If the Each is preceeded by a source c.t.Tap, it will be duplicated across branches to reduce
the number of step in the Flow.

Fixed bug in c.f.MultiMapReducePlanner where too many temp tap instances were being inserted while normalizing
the flow sources.

Changed c.t.Fields to fail if given duplicate field names.

Changed behavior if Hadoop FileInputSplit is not used and property "map.input.file" is not set. If there is one
source, it will returned as the source for the mapper stack, otherwise an exception is thrown. Subsequently joins
and merges of non-file sources is not supported until a discriminator can be passed to the mapper.

Fixed bug in c.t.Tuple where NPE was thrown under certain compareTo operations.

Fixed bug that prevented CoGrouping or Merging on the same source even though it was one or more Groupings away.

0.7.0

Changes project structure, removed 'examples' sub-project.

Updated to support Hadoop 0.17.x. This version is not API compatible with any Hadoop version less than 0.17.0.

Added ability to stop all c.f.Flows executing within a c.c.Cascade instance via the stop() method.

Changed c.f.FlowConnector to only take a Map of properties. These properties are passed downstream to various
subsystems. This removes the Hadoop JobConf constructor, but it still can be passed as a property value. Also
properties will be pushed into a defaul JobConf, bypassing any direct JobConf coupling in applications.

Changed c.f.Flow to automatically register a shutdown hook killing remote jobs on vm exit.

Changed c.f.Flow.stop() to immediately stop all running jobs.

Changed c.o.Operation to an interface and introduced c.o.BaseOperation. This makes creating custom Operation types
more flexible and intuitive. c.o.Filter, c.o.Function, c.o.Aggregator, and c.o.Assertion now extend c.o.Operation.

Added c.p.c.OuterJoin, c.p.c.MixedJoin, c.p.c.LeftJoin, and c.p.c.RightJoin c.p.c.CoGrouper classes. They
compliment the default c.p.c.InnerJoin CoGrouper class.

Added support for passing an intermediateSchemeClass to the underlying planner to be used as the default c.s.Scheme
for intermediate c.t.Tap instances internal to a given c.f.Flow.

Fixed bug where c.p.Group is immediately followed by another c.p.Group (or their sub-classes) and fields could not
be resolved between them.

Added support for c.t.Tap instances implementing c.f.FlowListener. If implemented, they will automatically be
added to the Flow event listeners collection and will receive Flow events.

Fixed case where multiple source c.t.Tap instances return true for the containsFile method. Now verifies only one
Tap contains the file, and fails otherwise.

Changed c.s.TextLine to not set numSinkParts to 1 by default. Now uses the natural number of parts.

Changed MapReduce planner to force an intermediate file between branches with Hadoop incompatible source Taps
on joins/merges. If the taps are compatible (have same Scheme), all branches will be processed in same Mapper
before the c.p.Group.

Added merge capabilities in c.p.GroupBy. This allows multiple input branches to be grouped as if a single stream.

Fixed bug in c.t.TapCollector where writing to a Sequence file threw a NPE.

Added c.f.MapReduceFlow to support custom MapReduce jobs, allowing them to participate in a Cascade job.

0.6.1

Changed thrown c.f.FlowException instances to include cause message.

Fixed bug where empty sink or source map was not detected.

0.6.0

Changed default argument selector for c.p.Every to be Fields.ALL, to be consistent with the default value of c.p.Each.

Added support for assembly traps. If an exception is thrown from inside an c.o.Operation, the offending Tuple
can be saved to a file for later processing, allowing the job to complete.

Added support for stream assertions. STRICT and VALID assertions can be built into a pipe assembly, and optionally
planned out during runtime. Assertions will throw exceptions if they fail.

Changed c.o.a.First, Last, Min, and Max to optionally ignore specified values. Useful if you do not wish
for a 'default' value to be considered first, or last in a set.

Changed c.o.a.Sum to take a Class for coercion of the result value.

Changes c.o.Max and Min to use infinity as initial values so zero is bigger than a really small number
for Max, and zero is smaller than a really big number for Min.

Changed order of JobConf initialization. c.f.FlowStep now is added to the JobConf last in order to catch
all lazily configured values.

Changed compile to include debug info by default.

Fixed bug in c.t.MultiTap where super scheme was not returned if available.

0.5.0

Added skipIfSinkExists property to c.f.Flow. Set to true if the c.c.Cascade should skip the Flow instance even
if the sink is stale and not set to be deleted on initialization.

Fixed bug in c.t.h.HttpFileSystem that URL escaped the ? prefixing the query string.

Fixed bug where a join with duplicate taps was not recognized during job planning. Now an appropriate error
message is displayed, instead of jobs completing with only one instance of the resource stream.

Fixed c.t.h.HttpFileSystem to remember authority information in the url and prefix it when missing.

Changed c.s.TextLine to accept either on or two source fields. If one, only the 'line' value
is sourced from the value, discarding the 'offset' value.

Added c.o.r.RegexSplitGenerator to support splitting single tuple values into multiple tuples based on a regex
delimiter. Includes new tests.

Added c.s.CascadeStats and c.s.FlowStats to provide access to current state and statistics of particular
Cascade, Flow, or the child Flows of a Cascade.

Added ability to sort grouping values with sort argument on c.p.GroupBy. Sorts can be reversed.

Added c.o.e.ExpressionFilter, the c.o.Filter analog to c.o.e.ExpressionFunction.

0.4.1

Fixed path normalization regex in c.u.Util where it munged any path starting with file:///.

0.4.0

Changed c.p.GroupBy default grouping fields to c.t.Fields.ALL from Fields.FIRST. This change provides a simple
way to sort a tuple stream based on the order of the tuple fields.

Changed c.f.FlowConnector to create c.f.Flow instances that will bypass the reducer if no c.p.Group is participating
in the assembly. Previoiusly Group instances were inserted if missing. This allows a chain of c.p.Every instances
to be used to process/filter a tuple stream without the invoking the reducer needlessly (if a sort isn't required).
This change also supports bypassing the default Hadoop OutputCollector in the mapper via the sink c.t.Tap instance.

Changed c.f.FlowStep behavior to run in 'local' mode if either the sink or source tap is a c.t.Lfs instance. This
allows for c.f.Flow instances to run mixed if configured to execute on a particular cluster by default. This behavior
supports complex import/export processes against the HDFS or other supported remote filesystem.

Changed behavior of c.t.Dfs to force use of HDFS. Previously Dfs would default to the local FileSystem
if the job was run in 'local'mode. Now a Dfs instance will cause failures if it cannot connect to a HDFS cluster.
Using c.t.Hfs will provide previous Dfs behavior. Hfs will use the 'default' filesystem if a scheme is not present
in the 'stringPath' (i.e. hdfs://host:port/some/path).

Added c.stats package to allow for collecting statics of Cascades, Flows, and FlowSteps.

Updated c.f.Flow and c.c.Cascade log messages to be easier to follow when executing many flow instances
simultaneously.

Added compression flag to c.s.TextLine. Can now toggle compression (Hadoop style compression) per Tap instance.
This prevents clusters with compression enabled by default to export text files with a .deflate extension.

Added support for bypassing Hadoop OutputCollector via Tap.setUseTapCollector() method. Setting to true will force
Cascading to use the c.t.TapCollector instead. This bypasses bugs in Hadoop with custom FileSystem types. This will
always be true for http(s) and s3tp filesystems when using a c.t.Hfs Tap type (atleast until HADOOP-3021 is resolved).

Added c.t.TupleCollector, complementing c.t.TupleIterator, for directly writing Tuple instances out via a c.t.Tap
instance.

Added c.f.FlowListener so that c.f.Flow instances can fire events on starting, completed, and throwable.

Changed c.t.h.S3HttpFileSystem so it can now create files remotely.

Renamed cascading.spill.threshold to cascading.cogroup.spill.threshold, so there is less a chance of collision.

Made numerous optimizations to improve overall performance. Namely split and merge of key/value tuples to remove
redundancy in the stream between the mapper and reducer.

Changed c.p.Operators to push c.o.Operation results directly through to next operation without intermediate
collection. This should improve pipelining of large result streams and lower runtime memory footprint.

Changed c.c.Cascade so it now runs Flows in parallel if Hadoop is clustered, and there are no dependencies between the
Flows.

Moved c.Cascade and related classed to c.cascade package. Wanted to preempt any future ugliness.

Added support in c.t.h.S3HttpFileSystem for these properties: fs.s3tp.awsAccessKeyId and fs.s3tp.awsSecretAccessKey

0.3.0

Added ability to push Log4j logger properties to mapper/reducer via JobConf.
Use jobConf.set("log4j.logger","logger1=LEVEL,logger2=LEVEL")

Added missing equals() and hashCode() in c.t.MultiTap.

Added c.t.h.ZipInputFormat (and ZipSplit) to support zip files. c.s.TextLine supports transparent
reading of zip files if the filename ends with .zip, but cannot write to them. This code is
loosely based on HADOOP-1824. If the underlying filesystem is hdfs or file, splits will be created
for each ZipEntry. Otherwise ZipEntries are iterated over to be more stream friendly. Progress status is
supported.

Added http, https, and s3tp read-only file systems to Hadoop. Use these URLs, respectively:
http://, https://, and s3tp://AWS_ACCESS_KEY_ID:AWS_SECRET_ACCESS_KEY@bucket-name/key

Added c.o.t.DateFormatter supporting text formatting of time stamps created by c.o.t.DateParser.

Fixed bug where in complex assemblies, some Scopes were not resolved.

Fixed bug where tap instances were not being inserted before some CoGroup joins if there was a previous Group in the
assembly.

Upgraded JGraphT to 0.7.3

Changed c.t.SpillableTupleList allows for iteration across entries.

Changed c.f.FlowException to optionally allow for printing of underlying pipe graph for debugging.

Added c.o.t.FieldFormatter function to format Tuples into complex strings using j.u.Formatter formatting.

Added c.o.a.Last aggregator to find the last value encountered in a group.

Changed c.o.a.Max and c.o.a.Min to maintain original value type. Will return null if no values are encountered.

Changed c.o.a.First to use Fields.ARG by default. Removed Fields constructor.

Added c.t.Fields.join(Fields...) method to allow for joining multiple Fields instances into a new instance.

Can retrieve Tuple values by field name through the TupleEntry class via the get(String) method.

Added c.t.TupleCollector interface to simplify the operation interfaces.

Added a Debug filter that will print to either stderr or stdout. Useful for debugging stream transformations.

Added CascadingTestCase base test class

Added Insert Function that allows for literal values to be inserted into the Tuple stream.

0.2.0

CoGroup will now spill to disk on extremely large co-groupings. Configurable via "cascading.spill.threshold".
Defaults to 10k elements.

java.util.Properties instances can be used to set defauls for FlowConnectors.

Fix for InnerJoin, the default join for CoGroup.

Introduced MultiTap to support concatenation of files into a pipe assembly.

RegexParser now fails on a failed match. Prevents it being used or behaving as a filter.

Fixed bug with PipeAssembly instances not properly being assimiliated into the pipeGraph.

Fixed assertion error thrown by JGraphT.

Renamed Tap method deleteOnInit to deleteOnSinkInit.

0.1.0

First release.

Show details Hide details

Change log

r580 by cwensel on Dec 13, 2008 Diff

release 0.10.0

Go to:

Older revisions

r579 by cwensel on Dec 12, 2008 Diff

version 0.10.0

r574 by cwensel on Dec 12, 2008 Diff

renamed repeat parameter to
numSelfJoins. value now relfects the
number of self joins to be performed

r568 by cwensel on Dec 11, 2008 Diff

Merge branch 'working'

All revisions of this file

File info

Size: 25629 bytes, 493 lines

View raw file