src/java/com/twitter/search/earlybird/archive/ArchiveTimeSlicer.java - main - githubmirror/the-algorithm

src/java/com/twitter/search/earlybird/archive/ArchiveTimeSlicer.java
322 строки12 KB
twitter-team
Twitter Recommendation Algorithm
01 апр 2023, 01:36
01 апр 2023, 01:36ef4c5eb
package com.twitter.search.earlybird.archive;

import java.io.IOException;
import java.util.ArrayList;
import java.util.Calendar;
import java.util.Collections;
import java.util.Comparator;
import java.util.Date;
import java.util.List;

import com.google.common.annotations.VisibleForTesting;
import com.google.common.base.Preconditions;
import com.google.common.base.Predicate;
import com.google.common.collect.Lists;


import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

import com.twitter.search.common.schema.thriftjava.ThriftIndexingEvent;
import com.twitter.search.common.util.io.MergingSortedRecordReader;
import com.twitter.search.common.util.io.recordreader.RecordReader;
import com.twitter.search.earlybird.config.TierConfig;
import com.twitter.search.earlybird.document.DocumentFactory;
import com.twitter.search.earlybird.document.ThriftIndexingEventDocumentFactory;
import com.twitter.search.earlybird.document.TweetDocument;


/**
 * Responsible for taking a number of daily status batches and partitioning them into time slices
 * which will be used to build segments.
 *
 * We try to put at most N number of tweets into a time slice.
 */
public class ArchiveTimeSlicer {
  private static final Logger LOG = LoggerFactory.getLogger(ArchiveTimeSlicer.class);

  private static final Comparator<TweetDocument> ASCENDING =
      (o1, o2) -> Long.compare(o1.getTweetID(), o2.getTweetID());

  private static final Comparator<TweetDocument> DESCENDING =
      (o1, o2) -> Long.compare(o2.getTweetID(), o1.getTweetID());

  // Represents a number of daily batches which will go into a segment.
  public static final class ArchiveTimeSlice {
    private Date startDate;
    private Date endDate;
    private int statusCount;
    private final DailyStatusBatches directory;
    private final ArchiveEarlybirdIndexConfig earlybirdIndexConfig;

    // This list is always ordered from oldest day, to the newest day.
    // For the on-disk archive, we reverse the days in getTweetReaders().
    private final List<DailyStatusBatch> batches = Lists.newArrayList();

    private ArchiveTimeSlice(DailyStatusBatches directory,
                             ArchiveEarlybirdIndexConfig earlybirdIndexConfig) {
      this.directory = directory;
      this.earlybirdIndexConfig = earlybirdIndexConfig;
    }

    public Date getEndDate() {
      return endDate;
    }

    public int getStatusCount() {
      return statusCount;
    }

    public int getNumHashPartitions() {
      return batches.isEmpty() ? 0 : batches.get(0).getNumHashPartitions();
    }

    /**
     * Returns a reader for reading tweets from this timeslice.
     *
     * @param archiveSegment The segment to which the timeslice belongs.
     * @param documentFactory The ThriftIndexingEvent to TweetDocument converter.
     * @param filter A filter that determines what dates should be read.
     */
    public RecordReader<TweetDocument> getStatusReader(
        ArchiveSegment archiveSegment,
        DocumentFactory<ThriftIndexingEvent> documentFactory,
        Predicate<Date> filter) throws IOException {
      // We no longer support ThriftStatus based document factories.
      Preconditions.checkState(documentFactory instanceof ThriftIndexingEventDocumentFactory);

      final int hashPartitionID = archiveSegment.getHashPartitionID();
      List<RecordReader<TweetDocument>> readers = new ArrayList<>(batches.size());
      List<DailyStatusBatch> orderedForReading = orderBatchesForReading(batches);
      LOG.info("Creating new status reader for hashPartition: "
          + hashPartitionID + " timeslice: " + getDescription());

      for (DailyStatusBatch batch : orderedForReading) {
        if (filter.apply(batch.getDate())) {
          LOG.info("Adding reader for " + batch.getDate() + " " + getDescription());
          PartitionedBatch partitionedBatch = batch.getPartition(hashPartitionID);
          // Don't even try to create a reader if the partition is empty.
          // There does not seem to be any problem in production now, but HDFS FileSystem's javadoc
          // does indicate that listStatus() is allowed to throw a FileNotFoundException if the
          // partition does not exist. This check makes the code more robust against future
          // HDFS FileSystem implementation changes.
          if (partitionedBatch.getStatusCount() > 0) {
            RecordReader<TweetDocument> tweetReaders = partitionedBatch.getTweetReaders(
                archiveSegment,
                directory.getStatusPathToUseForDay(batch.getDate()),
                documentFactory);
            readers.add(tweetReaders);
          }
        } else {
          LOG.info("Filtered reader for " + batch.getDate() + " " + getDescription());
        }
      }

      LOG.info("Creating reader for timeslice: " + getDescription()
          + " with " + readers.size() + " readers");

      return new MergingSortedRecordReader<TweetDocument>(getMergingComparator(), readers);
    }

    private List<DailyStatusBatch> orderBatchesForReading(List<DailyStatusBatch> orderedBatches) {
      // For the index formats using stock lucene, we want the most recent days to be indexed first.
      // In the twitter in-memory optimized indexes, older tweets will be added first, and
      // optimization will reverse the documents to make most recent tweets be first.
      return this.earlybirdIndexConfig.isUsingLIFODocumentOrdering()
          ? orderedBatches : Lists.reverse(orderedBatches);
    }

    private Comparator<TweetDocument> getMergingComparator() {
      // We always want to retrieve larger tweet ids first.
      // LIFO means that the smaller ids get inserted first --> ASCENDING order.
      // FIFO would mean that we want to first insert the larger ids --> DESCENDING order.
      return this.earlybirdIndexConfig.isUsingLIFODocumentOrdering()
          ? ASCENDING : DESCENDING;
    }

    /**
     * Returns the smallest indexed tweet ID in this timeslice for the given partition.
     *
     * @param hashPartitionID The partition.
     */
    public long getMinStatusID(int hashPartitionID) {
      if (batches.isEmpty()) {
        return 0;
      }

      for (int i = 0; i < batches.size(); i++) {
        long minStatusID = batches.get(i).getPartition(hashPartitionID).getMinStatusID();
        if (minStatusID != DailyStatusBatch.EMPTY_BATCH_STATUS_ID) {
          return minStatusID;
        }
      }

      return 0;
    }

    /**
     * Returns the highest indexed tweet ID in this timeslice for the given partition.
     *
     * @param hashPartitionID The partition.
     */
    public long getMaxStatusID(int hashPartitionID) {
      if (batches.isEmpty()) {
        return Long.MAX_VALUE;
      }

      for (int i = batches.size() - 1; i >= 0; i--) {
        long maxStatusID = batches.get(i).getPartition(hashPartitionID).getMaxStatusID();
        if (maxStatusID != DailyStatusBatch.EMPTY_BATCH_STATUS_ID) {
          return maxStatusID;
        }
      }

      return Long.MAX_VALUE;
    }

    /**
     * Returns a string with some information for this timeslice.
     */
    public String getDescription() {
      StringBuilder builder = new StringBuilder();
      builder.append("TimeSlice[start date=");
      builder.append(DailyStatusBatches.DATE_FORMAT.format(startDate));
      builder.append(", end date=");
      builder.append(DailyStatusBatches.DATE_FORMAT.format(endDate));
      builder.append(", status count=");
      builder.append(statusCount);
      builder.append(", days count=");
      builder.append(batches.size());
      builder.append("]");
      return builder.toString();
    }
  }

  private final int maxSegmentSize;
  private final DailyStatusBatches dailyStatusBatches;
  private final Date tierStartDate;
  private final Date tierEndDate;
  private final ArchiveEarlybirdIndexConfig earlybirdIndexConfig;

  private List<ArchiveTimeSlice> lastCachedTimeslices = null;

  public ArchiveTimeSlicer(int maxSegmentSize,
                           DailyStatusBatches dailyStatusBatches,
                           ArchiveEarlybirdIndexConfig earlybirdIndexConfig) {
    this(maxSegmentSize, dailyStatusBatches, TierConfig.DEFAULT_TIER_START_DATE,
        TierConfig.DEFAULT_TIER_END_DATE, earlybirdIndexConfig);
  }

  public ArchiveTimeSlicer(int maxSegmentSize,
                           DailyStatusBatches dailyStatusBatches,
                           Date tierStartDate,
                           Date tierEndDate,
                           ArchiveEarlybirdIndexConfig earlybirdIndexConfig) {
    this.maxSegmentSize = maxSegmentSize;
    this.dailyStatusBatches = dailyStatusBatches;
    this.tierStartDate = tierStartDate;
    this.tierEndDate = tierEndDate;
    this.earlybirdIndexConfig = earlybirdIndexConfig;
  }

  private boolean cacheIsValid() throws IOException {
    return lastCachedTimeslices != null
        && !lastCachedTimeslices.isEmpty()
        && cacheIsValid(lastCachedTimeslices.get(lastCachedTimeslices.size() - 1).endDate);
  }

  private boolean cacheIsValid(Date lastDate) throws IOException {
    if (lastCachedTimeslices == null || lastCachedTimeslices.isEmpty()) {
      return false;
    }

    // Check if we have a daily batch newer than the last batch used for the newest timeslice.
    Calendar cal = Calendar.getInstance();
    cal.setTime(lastDate);
    cal.add(Calendar.DATE, 1);
    Date nextDate = cal.getTime();

    boolean foundBatch = dailyStatusBatches.hasValidBatchForDay(nextDate);

    LOG.info("Checking cache: Looked for valid batch for day {}. Found: {}",
        DailyStatusBatches.DATE_FORMAT.format(nextDate), foundBatch);

    return !foundBatch;
  }

  private boolean timesliceIsFull(ArchiveTimeSlice timeSlice, DailyStatusBatch batch) {
    return timeSlice.statusCount + batch.getMaxPerPartitionStatusCount() > maxSegmentSize;
  }

  private void doTimeSlicing() throws IOException {
    dailyStatusBatches.refresh();

    lastCachedTimeslices = Lists.newArrayList();
    ArchiveTimeSlice currentTimeSlice = null;

    // Iterate over each day and add it to the current timeslice, until it gets full.
    for (DailyStatusBatch batch : dailyStatusBatches.getStatusBatches()) {
      if (!batch.isValid()) {
        LOG.warn("Skipping hole: " + batch.getDate());
        continue;
      }

      if (currentTimeSlice == null || timesliceIsFull(currentTimeSlice, batch)) {
        if (currentTimeSlice != null) {
          LOG.info("Filled timeslice: " + currentTimeSlice.getDescription());
        }
        currentTimeSlice = new ArchiveTimeSlice(dailyStatusBatches, earlybirdIndexConfig);
        currentTimeSlice.startDate = batch.getDate();
        lastCachedTimeslices.add(currentTimeSlice);
      }

      currentTimeSlice.endDate = batch.getDate();
      currentTimeSlice.statusCount += batch.getMaxPerPartitionStatusCount();
      currentTimeSlice.batches.add(batch);
    }
    LOG.info("Last timeslice: {}", currentTimeSlice.getDescription());

    LOG.info("Done with time slicing. Number of timeslices: {}",
        lastCachedTimeslices.size());
  }

  /**
   * Returns all timeslices for this earlybird.
   */
  public List<ArchiveTimeSlice> getTimeSlices() throws IOException {
    if (cacheIsValid()) {
      return lastCachedTimeslices;
    }

    LOG.info("Cache is outdated. Loading new daily batches now...");

    doTimeSlicing();

    return lastCachedTimeslices != null ? Collections.unmodifiableList(lastCachedTimeslices) : null;
  }

  /**
   * Return the timeslices that overlap the tier start/end date ranges if they are specified
   */
  public List<ArchiveTimeSlice> getTimeSlicesInTierRange() throws IOException {
    List<ArchiveTimeSlice> timeSlices = getTimeSlices();
    if (tierStartDate == TierConfig.DEFAULT_TIER_START_DATE
        && tierEndDate == TierConfig.DEFAULT_TIER_END_DATE) {
      return timeSlices;
    }

    List<ArchiveTimeSlice> filteredTimeSlice = Lists.newArrayList();
    for (ArchiveTimeSlice timeSlice : timeSlices) {
      if (timeSlice.startDate.before(tierEndDate) && !timeSlice.endDate.before(tierStartDate)) {
        filteredTimeSlice.add(timeSlice);
      }
    }

    return filteredTimeSlice;
  }

  @VisibleForTesting
  protected DailyStatusBatches getDailyStatusBatches() {
    return dailyStatusBatches;
  }
}