@Experimental(value=SOURCE_SINK) public abstract static class BoundedSource.BoundedReader<T> extends Source.Reader<T>
Reader
that reads a bounded amount of input and supports some additional
operations, such as progress estimation and dynamic work rebalancing.
Once Source.Reader.start()
or Source.Reader.advance()
has returned false, neither will be called
again on this object.
All methods will be run from the same thread except splitAtFraction(double)
,
getFractionConsumed()
, getCurrentSource()
, getSplitPointsConsumed()
,
and getSplitPointsRemaining()
, all of which can be called concurrently
from a different thread. There will not be multiple concurrent calls to
splitAtFraction(double)
.
It must be safe to call splitAtFraction(double)
, getFractionConsumed()
,
getCurrentSource()
, getSplitPointsConsumed()
, and
getSplitPointsRemaining()
concurrently with other methods.
Additionally, a successful splitAtFraction(double)
call must, by definition, cause
getCurrentSource()
to start returning a different value.
Callers of getCurrentSource()
need to be aware of the possibility that the returned
value can change at any time, and must only access the properties of the source returned by
getCurrentSource()
which do not change between splitAtFraction(double)
calls.
splitAtFraction(double)
In the course of dynamic work rebalancing, the method splitAtFraction(double)
may be called concurrently with Source.Reader.advance()
or Source.Reader.start()
. It is critical that
their interaction is implemented in a thread-safe way, otherwise data loss is possible.
Sources which support dynamic work rebalancing should use
RangeTracker
to manage the (source-specific)
range of positions that is being split.
Modifier and Type | Field and Description |
---|---|
static long |
SPLIT_POINTS_UNKNOWN
A constant to use as the return value for
getSplitPointsConsumed() or
getSplitPointsRemaining() when the exact value is unknown. |
Constructor and Description |
---|
BoundedReader() |
Modifier and Type | Method and Description |
---|---|
abstract BoundedSource<T> |
getCurrentSource()
Returns a
Source describing the same input that this Reader currently reads
(including items already read). |
Instant |
getCurrentTimestamp()
By default, returns the minimum possible timestamp.
|
java.lang.Double |
getFractionConsumed()
Returns a value in [0, 1] representing approximately what fraction of the
current source this reader has read so far, or null if such
an estimate is not available. |
long |
getSplitPointsConsumed()
Returns the total amount of parallelism in the consumed (returned and processed) range of
this reader's current
BoundedSource (as would be returned by
getCurrentSource() ). |
long |
getSplitPointsRemaining()
Returns the total amount of parallelism in the unprocessed part of this reader's current
BoundedSource (as would be returned by getCurrentSource() ). |
BoundedSource<T> |
splitAtFraction(double fraction)
Tells the reader to narrow the range of the input it's going to read and give up
the remainder, so that the new range would contain approximately the given
fraction of the amount of data in the current range.
|
advance, close, getCurrent, start
public static final long SPLIT_POINTS_UNKNOWN
getSplitPointsConsumed()
or
getSplitPointsRemaining()
when the exact value is unknown.@Nullable public java.lang.Double getFractionConsumed()
current source
this reader has read so far, or null
if such
an estimate is not available.
It is recommended that this method should satisfy the following properties:
Source.Reader.start()
call.
Source.Reader.start()
or Source.Reader.advance()
call that returns false.
By default, returns null to indicate that this cannot be estimated.
splitAtFraction(double)
is implemented, this method can be called concurrently to other
methods (including itself), and it is therefore critical for it to be implemented
in a thread-safe way.public long getSplitPointsConsumed()
BoundedSource
(as would be returned by
getCurrentSource()
). This corresponds to all split point records (see
RangeTracker
) returned by this reader, excluding the last split point
returned if the reader is not finished.
Consider the following examples: (1) An input that can be read in parallel down to the
individual records, such as CountingSource.upTo(long)
, is called "perfectly splittable".
(2) a "block-compressed" file format such as AvroIO
, in which a block of records has
to be read as a whole, but different blocks can be read in parallel. (3) An "unsplittable"
input such as a cursor in a database.
reader
that is unstarted (aka, has never had a call to
Source.Reader.start()
) has a consumed parallelism of 0. This condition holds independent of whether
the input is splittable.
reader
that has only returned its first element (aka,
has never had a call to Source.Reader.advance()
) has a consumed parallelism of 0: the first element
is the current element and is still being processed. This condition holds independent of
whether the input is splittable.
Source.Reader.start()
returned false), the
consumed parallelism is 0. This condition holds independent of whether the input is
splittable.
Source.Reader.start()
returned true and
a call to Source.Reader.advance()
has returned false), the value returned must be at least 1
and should equal the total parallelism in the source.
Source.Reader.advance()
returned false, at
which point it becomes 1.
A reader that is implemented using a RangeTracker
is encouraged to use the
range tracker's ability to count split points to implement this method. See
OffsetBasedSource.OffsetBasedReader
and OffsetRangeTracker
for an example.
Defaults to SPLIT_POINTS_UNKNOWN
. Any value less than 0 will be interpreted
as unknown.
BoundedSource.BoundedReader
for information about thread safety.getSplitPointsRemaining()
public long getSplitPointsRemaining()
BoundedSource
(as would be returned by getCurrentSource()
). This corresponds
to all unprocessed split point records (see RangeTracker
), including the last
split point returned, in the remainder part of the source.
This function should be implemented only in addition to
getSplitPointsConsumed()
and only if an exact value can be
returned.
Consider the following examples: (1) An input that can be read in parallel down to the
individual records, such as CountingSource.upTo(long)
, is called "perfectly splittable".
(2) a "block-compressed" file format such as AvroIO
, in which a block of records has
to be read as a whole, but different blocks can be read in parallel. (3) An "unsplittable"
input such as a cursor in a database.
Assume for examples (1) and (2) that the number of records or blocks remaining is known:
reader
for which the last call to Source.Reader.start()
or
Source.Reader.advance()
has returned true should should not return 0, because this reader itself
represents parallelism at least 1. This condition holds independent of whether the input is
splittable.
Source.Reader.start()
or Source.Reader.advance()
) has returned false
should return a value of 0. This condition holds independent of whether the input is
splittable.
Defaults to SPLIT_POINTS_UNKNOWN
. Any value less than 0 will be interpreted as
unknown.
BoundedSource.BoundedReader
for information about thread safety.getSplitPointsConsumed()
public abstract BoundedSource<T> getCurrentSource()
Source
describing the same input that this Reader
currently reads
(including items already read).
Reader subclasses can use this method for convenience to access unchanging properties of the source being read. Alternatively, they can cache these properties in the constructor.
The framework will call this method in the course of dynamic work rebalancing, e.g. after
a successful splitAtFraction(double)
call.
Remember that Source
objects must always be immutable. However, the return value
of this function may be affected by dynamic work rebalancing, happening asynchronously via
splitAtFraction(double)
, meaning it can return a different
Source
object. However, the returned object itself will still itself be immutable.
Callers must take care not to rely on properties of the returned source that may be
asynchronously changed as a result of this process (e.g. do not cache an end offset when
reading a file).
For convenience, subclasses should usually return the most concrete subclass of
Source
possible.
In practice, the implementation of this method should nearly always be one of the following:
getCurrentSource()
: delegate to base class. In this case, it is almost always
an error for the subclass to maintain its own copy of the source.
public FooReader(FooSource<T> source) {
super(source);
}
public FooSource<T> getCurrentSource() {
return (FooSource<T>)super.getCurrentSource();
}
private final FooSource<T> source;
public FooReader(FooSource<T> source) {
this.source = source;
}
public FooSource<T> getCurrentSource() {
return source;
}
BoundedSource.BoundedReader
that explicitly supports dynamic work rebalancing:
maintain a variable pointing to an immutable source object, and protect it with
synchronization.
private FooSource<T> source;
public FooReader(FooSource<T> source) {
this.source = source;
}
public synchronized FooSource<T> getCurrentSource() {
return source;
}
public synchronized FooSource<T> splitAtFraction(double fraction) {
...
FooSource<T> primary = ...;
FooSource<T> residual = ...;
this.source = primary;
return residual;
}
getCurrentSource
in class Source.Reader<T>
@Nullable public BoundedSource<T> splitAtFraction(double fraction)
Returns a BoundedSource
representing the remainder.
BoundedSource<T> initial = reader.getCurrentSource();
BoundedSource<T> residual = reader.splitAtFraction(fraction);
BoundedSource<T> primary = reader.getCurrentSource();
This method should return null
if the split cannot be performed for this fraction
while satisfying the semantics above. E.g., a reader that reads a range of offsets
in a file should return null
if it is already past the position in its range
corresponding to the given fraction. In this case, the method MUST have no effect
(the reader must behave as if the method hadn't been called at all).
It is also very important that this method always completes quickly. In particular, it should not perform or wait on any blocking operations such as I/O, RPCs etc. Violating this requirement may stall completion of the work item or even cause it to fail.
It is incorrect to make both this method and Source.Reader.start()
/Source.Reader.advance()
synchronized
, because those methods can perform blocking operations, and then
this method would have to wait for those calls to complete.
RangeTracker
makes it easy to implement
this method safely and correctly.
By default, returns null to indicate that splitting is not possible.
public Instant getCurrentTimestamp() throws java.util.NoSuchElementException
getCurrentTimestamp
in class Source.Reader<T>
java.util.NoSuchElementException
- if the reader is at the beginning of the input and
Source.Reader.start()
or Source.Reader.advance()
wasn't called, or if the last Source.Reader.start()
or
Source.Reader.advance()
returned false
.