io.deephaven.engine.table.impl.updateby.UpdateByWindowRollingTime Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of deephaven-engine-table Show documentation
Show all versions of deephaven-engine-table Show documentation
Engine Table: Implementation and closely-coupled utilities
package io.deephaven.engine.table.impl.updateby;
import io.deephaven.base.ringbuffer.LongRingBuffer;
import io.deephaven.base.verify.Assert;
import io.deephaven.chunk.LongChunk;
import io.deephaven.chunk.WritableIntChunk;
import io.deephaven.chunk.attributes.Values;
import io.deephaven.engine.rowset.*;
import io.deephaven.engine.table.ChunkSource;
import io.deephaven.engine.table.ColumnSource;
import io.deephaven.engine.table.TableUpdate;
import io.deephaven.engine.table.impl.ssa.LongSegmentedSortedArray;
import io.deephaven.engine.table.iterators.ChunkedLongColumnIterator;
import io.deephaven.engine.table.iterators.LongColumnIterator;
import org.jetbrains.annotations.NotNull;
import org.jetbrains.annotations.Nullable;
import java.util.BitSet;
import java.util.stream.IntStream;
import static io.deephaven.util.QueryConstants.NULL_INT;
import static io.deephaven.util.QueryConstants.NULL_LONG;
/**
* This is the specialization of {@link UpdateByWindow} that handles time-based `windowed` operators. These operators
* maintain a window of data based on a timestamp column rather than row distances. Window-based operators must maintain
* a buffer of `influencer` values to add to the rolling window as the current row changes.
*/
class UpdateByWindowRollingTime extends UpdateByWindowRollingBase {
private static final int RING_BUFFER_INITIAL_SIZE = 128;
public static class UpdateByWindowTimeBucketContext extends UpdateByWindowRollingBucketContext {
public UpdateByWindowTimeBucketContext(final TrackingRowSet sourceRowSet,
@NotNull final ColumnSource timestampColumnSource,
@Nullable final LongSegmentedSortedArray timestampSsa,
final TrackingRowSet timestampValidRowSet,
final boolean timestampsModified,
final int chunkSize,
final boolean initialStep) {
super(sourceRowSet, timestampColumnSource, timestampSsa, timestampValidRowSet, timestampsModified,
chunkSize, initialStep);
}
}
UpdateByWindowRollingTime(
UpdateByOperator[] operators,
int[][] operatorSourceSlots,
@NotNull String timestampColumnName,
long prevUnits,
long fwdUnits) {
super(operators, operatorSourceSlots, prevUnits, fwdUnits, timestampColumnName);
}
@Override
UpdateByWindow copy() {
final UpdateByOperator[] copiedOperators = new UpdateByOperator[this.operators.length];
for (int ii = 0; ii < copiedOperators.length; ii++) {
copiedOperators[ii] = this.operators[ii].copy();
}
return new UpdateByWindowRollingTime(copiedOperators, operatorInputSourceSlots, timestampColumnName, prevUnits,
fwdUnits);
}
@Override
public UpdateByWindowBucketContext makeWindowContext(final TrackingRowSet sourceRowSet,
final ColumnSource timestampColumnSource,
final LongSegmentedSortedArray timestampSsa,
final TrackingRowSet timestampValidRowSet,
final boolean timestampsModified,
final int chunkSize,
final boolean isInitializeStep) {
return new UpdateByWindowTimeBucketContext(sourceRowSet, timestampColumnSource, timestampSsa,
timestampValidRowSet, timestampsModified, chunkSize, isInitializeStep);
}
/**
* Finding the `affected` and `influencer` rowsets for a windowed operation is complex. We must identify modified
* rows (including added rows) and deleted rows and determine which rows are `affected` by the change given the
* window parameters. After these rows have been identified, must determine which rows will be needed to recompute
* these values (i.e. that fall within the window and will `influence` this computation).
*/
private static WritableRowSet computeAffectedRowsTime(
final UpdateByWindowTimeBucketContext ctx,
final ChunkSource.GetContext tsContext,
final RowSet subset, long revNanos, long fwdNanos, boolean usePrev) {
// swap fwd/rev to get the affected windows
return computeInfluencerRowsTime(ctx, tsContext, subset, fwdNanos, revNanos, usePrev);
}
private static WritableRowSet computeInfluencerRowsTime(
final UpdateByWindowTimeBucketContext ctx,
final ChunkSource.GetContext tsContext,
final RowSet subset,
long revNanos, long fwdNanos, boolean usePrev) {
try (final RowSequence.Iterator it = subset.getRowSequenceIterator()) {
final RowSetBuilderSequential builder = RowSetFactory.builderSequential();
LongSegmentedSortedArray.Iterator ssaIt = ctx.timestampSsa.iterator(false, false);
while (it.hasMore() && ssaIt.hasNext()) {
final RowSequence rs = it.getNextRowSequenceWithLength(ctx.workingChunkSize);
final int rsSize = rs.intSize();
LongChunk timestamps = usePrev
? ctx.timestampColumnSource.getPrevChunk(tsContext, rs).asLongChunk()
: ctx.timestampColumnSource.getChunk(tsContext, rs).asLongChunk();
for (int ii = 0; ii < rsSize; ii++) {
final long ts = timestamps.get(ii);
if (ts == NULL_LONG) {
// if the timestamp is null, the row won't belong to any set and we can pretend it doesn't exist
continue;
}
// look at every row timestamp, compute the head and tail in nanos
final long head = ts - revNanos;
final long tail = ts + fwdNanos;
// advance the iterator to the beginning of the window
if (ssaIt.nextValue() < head) {
ssaIt.advanceToBeforeFirst(head);
if (!ssaIt.hasNext()) {
// SSA is exhausted
break;
}
}
Assert.eqTrue(ssaIt.hasNext() && ssaIt.nextValue() >= head,
"SSA Iterator outside of window");
// step through the SSA and collect keys until outside the window
while (ssaIt.hasNext() && ssaIt.nextValue() <= tail) {
builder.appendKey(ssaIt.nextKey());
ssaIt.next();
}
if (!ssaIt.hasNext()) {
// SSA is exhausted
break;
}
}
}
return builder.build();
}
}
/**
* Windowed by time/ticks is complex to compute: must find all the changed rows and rows that would be affected by
* the changes (includes newly added rows) and need to be recomputed. Then include all the rows that are affected by
* deletions (if any). After the affected rows have been identified, determine which rows will be needed to compute
* new values for the affected rows (influencer rows)
*/
@Override
public void computeAffectedRowsAndOperators(UpdateByWindowBucketContext context, @NotNull TableUpdate upstream) {
UpdateByWindowTimeBucketContext ctx = (UpdateByWindowTimeBucketContext) context;
try (final ChunkSource.GetContext tsContext = ctx.timestampColumnSource.makeGetContext(ctx.workingChunkSize)) {
if (upstream.empty() || ctx.sourceRowSet.isEmpty()) {
// No further work will be done on this context
finalizeWindowBucket(context);
return;
}
// all rows are affected on the initial step
if (ctx.initialStep) {
ctx.affectedRows = ctx.sourceRowSet;
// Get the exact set of rows needed to compute the initial row set
ctx.influencerRows =
computeInfluencerRowsTime(ctx, tsContext, ctx.affectedRows, prevUnits, fwdUnits, false);
// mark all operators as affected by this update
context.dirtyOperatorIndices = IntStream.range(0, operators.length).toArray();
context.dirtyOperators = new BitSet(operators.length);
context.dirtyOperators.set(0, operators.length);
ctx.isDirty = true;
return;
}
// determine which operators are affected by this update
processUpdateForContext(context, upstream);
if (!ctx.isDirty) {
// No further work will be done on this context
finalizeWindowBucket(context);
return;
}
final WritableRowSet tmpAffected = RowSetFactory.empty();
// consider the modifications only when input or timestamp columns were modified
if (upstream.modified().isNonempty() && (ctx.timestampsModified || ctx.inputModified)) {
// recompute all windows that have the modified rows in their window
try (final RowSet modifiedAffected =
computeAffectedRowsTime(ctx, tsContext, upstream.modified(), prevUnits, fwdUnits, false)) {
tmpAffected.insert(modifiedAffected);
}
if (ctx.timestampsModified) {
// recompute all windows previously containing the modified rows
// after the timestamp modifications
try (final WritableRowSet modifiedAffectedPrev =
computeAffectedRowsTime(ctx, tsContext, upstream.getModifiedPreShift(), prevUnits, fwdUnits,
true)) {
// we used the SSA (post-shift) to get these keys, no need to shift
// retain only the rows that still exist in the sourceRowSet
modifiedAffectedPrev.retain(ctx.timestampValidRowSet);
tmpAffected.insert(modifiedAffectedPrev);
}
// re-compute all modified rows, they have new windows after the timestamp modifications
tmpAffected.insert(upstream.modified());
}
}
if (upstream.added().isNonempty()) {
// add the new rows and any cascading changes from inserting rows
final long prev = Math.max(0, prevUnits);
final long fwd = Math.max(0, fwdUnits);
try (final RowSet addedAffected =
computeAffectedRowsTime(ctx, tsContext, upstream.added(), prev, fwd, false)) {
tmpAffected.insert(addedAffected);
}
// compute all new rows
tmpAffected.insert(upstream.added());
}
// other rows can be affected by removes
if (upstream.removed().isNonempty()) {
final long prev = Math.max(0, prevUnits);
final long fwd = Math.max(0, fwdUnits);
try (final WritableRowSet removedAffected =
computeAffectedRowsTime(ctx, tsContext, upstream.removed(), prev, fwd, true)) {
// we used the SSA (post-shift) to get these keys, no need to shift
// retain only the rows that still exist in the sourceRowSet
removedAffected.retain(ctx.timestampValidRowSet);
tmpAffected.insert(removedAffected);
}
}
ctx.affectedRows = tmpAffected;
if (ctx.affectedRows.isEmpty()) {
// No further work will be done on this context
finalizeWindowBucket(context);
ctx.isDirty = false;
return;
}
// now get influencer rows for the affected rows
ctx.influencerRows =
computeInfluencerRowsTime(ctx, tsContext, ctx.affectedRows, prevUnits, fwdUnits, false);
}
}
private long nextLongOrMax(LongColumnIterator it) {
return it.hasNext() ? it.nextLong() : Long.MAX_VALUE;
}
public void computeWindows(UpdateByWindowRollingBucketContext context) {
UpdateByWindowTimeBucketContext ctx = (UpdateByWindowTimeBucketContext) context;
try (final RowSequence.Iterator affectedRowsIt = ctx.affectedRows.getRowSequenceIterator();
final ChunkSource.GetContext affectedTsGetContext =
ctx.timestampColumnSource.makeGetContext(ctx.workingChunkSize);
final LongColumnIterator influencerTsTailIt =
new ChunkedLongColumnIterator(ctx.timestampColumnSource, ctx.influencerRows)) {
final LongRingBuffer timestampWindowBuffer = new LongRingBuffer(RING_BUFFER_INITIAL_SIZE);
int affectedChunkOffset = 0;
long influencerTs = nextLongOrMax(influencerTsTailIt);
while (affectedRowsIt.hasMore()) {
// NOTE: We did not put null values into our SSA and our influencer rowset is built using the
// SSA. There should be no null timestamps considered in the rolling windows
final RowSequence affectedRs = affectedRowsIt.getNextRowSequenceWithLength(ctx.workingChunkSize);
final long chunkSize = affectedRs.intSize();
long totalPushCount = 0;
final LongChunk affectedTsChunk =
ctx.timestampColumnSource.getChunk(affectedTsGetContext, affectedRs).asLongChunk();
final WritableIntChunk pushChunk = ctx.pushChunks[affectedChunkOffset];
final WritableIntChunk popChunk = ctx.popChunks[affectedChunkOffset];
for (int ii = 0; ii < chunkSize; ii++) {
final long affectedTs = affectedTsChunk.get(ii);
if (affectedTs == NULL_LONG) {
// This signifies that does not belong to a time window.
popChunk.set(ii, NULL_INT);
pushChunk.set(ii, NULL_INT);
continue;
}
// Compute the head and tail timestamps (inclusive).
final long head = affectedTs - prevUnits;
final long tail = affectedTs + fwdUnits;
// Pop out all values from the current window that are not in the new window.
long popCount = 0;
while (!timestampWindowBuffer.isEmpty() && timestampWindowBuffer.front() < head) {
timestampWindowBuffer.remove();
popCount++;
}
// Push in all values that are in the new window (inclusive of tail).
long pushCount = 0;
while (influencerTs <= tail) {
// Add this value to the buffer before advancing.
timestampWindowBuffer.add(influencerTs);
influencerTs = nextLongOrMax(influencerTsTailIt);
pushCount++;
}
// Write the push and pop counts to the chunks.
popChunk.set(ii, Math.toIntExact(popCount));
pushChunk.set(ii, Math.toIntExact(pushCount));
totalPushCount += pushCount;
}
ctx.influencerCounts[affectedChunkOffset] = Math.toIntExact(totalPushCount);
ctx.maxGetContextSize = Math.max(ctx.maxGetContextSize, Math.toIntExact(totalPushCount));
affectedChunkOffset++;
}
}
}
}
© 2015 - 2024 Weber Informatics LLC | Privacy Policy