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

src/java/com/twitter/search/earlybird/archive/PartitionedBatch.java
333 строки13 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.FileNotFoundException;
import java.io.IOException;
import java.util.Comparator;
import java.util.Date;
import java.util.List;
import java.util.concurrent.TimeUnit;
import java.util.regex.Matcher;
import java.util.regex.Pattern;

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

import org.apache.commons.io.IOUtils;
import org.apache.hadoop.fs.FileStatus;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.fs.PathFilter;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

import com.twitter.search.common.config.Config;
import com.twitter.search.common.metrics.SearchCounter;
import com.twitter.search.common.partitioning.snowflakeparser.SnowflakeIdParser;
import com.twitter.search.common.schema.earlybird.EarlybirdThriftDocumentUtil;
import com.twitter.search.common.schema.thriftjava.ThriftIndexingEvent;
import com.twitter.search.common.util.date.DateUtil;
import com.twitter.search.common.util.io.EmptyRecordReader;
import com.twitter.search.common.util.io.LzoThriftBlockFileReader;
import com.twitter.search.common.util.io.MergingSortedRecordReader;
import com.twitter.search.common.util.io.TransformingRecordReader;
import com.twitter.search.common.util.io.recordreader.RecordReader;
import com.twitter.search.earlybird.common.config.EarlybirdConfig;
import com.twitter.search.earlybird.document.DocumentFactory;
import com.twitter.search.earlybird.document.TweetDocument;
import com.twitter.search.earlybird.partition.HdfsUtil;

/**
 * A batch of pre-processed tweets for a single hash partition from a particular day.
 */
public class PartitionedBatch {
  private static final Logger LOG = LoggerFactory.getLogger(PartitionedBatch.class);
  private static final Date START_DATE_INCLUSIVE = DateUtil.toDate(2006, 03, 21);
  private static final String STATUS_COUNT_FILE_PREFIX = "_status_count_";
  private static final Pattern STATUS_COUNT_FILE_PATTERN =
      Pattern.compile(STATUS_COUNT_FILE_PREFIX + "(\\d+)_minid_(\\d+)_maxid_(\\d+)");
  private static final int MAXIMUM_OUT_OF_ORDER_TOLERANCE_HOURS =
      EarlybirdConfig.getInt("archive_max_out_of_order_tolerance_hours", 12);
  private static final int READER_INIT_IOEXCEPTION_RETRIES = 20;
  private static final PathFilter LZO_DATA_FILES_FILTER = file -> file.getName().endsWith(".lzo");
  private static final PathFilter TXT_DATA_FILES_FILTER = file -> file.getName().endsWith(".txt");

  private static final Comparator<ThriftIndexingEvent> DESC_THRIFT_INDEXING_EVENT_COMPARATOR =
      (o1, o2) -> ComparisonChain.start()
          .compare(o2.getSortId(), o1.getSortId())
          .compare(o2.getUid(), o1.getUid())
          .result();

  // Number archive tweets skipped because they are too out-of-order.
  private static final SearchCounter OUT_OF_ORDER_STATUSES_SKIPPED =
      SearchCounter.export("out_of_order_archive_statuses_skipped");

  @VisibleForTesting
  protected static final long MAXIMUM_OUT_OF_ORDER_TOLERANCE_MILLIS =
      TimeUnit.HOURS.toMillis(MAXIMUM_OUT_OF_ORDER_TOLERANCE_HOURS);

  private final Date date;
  private final Path path;
  private int statusCount;
  private long minStatusID;
  private long maxStatusID;
  private final int hashPartitionID;
  private boolean hasStatusCountFile;
  private final int numHashPartitions;

  @VisibleForTesting
  public PartitionedBatch(
      Path path,
      int hashPartitionID,
      int numHashPartitions,
      Date date) {
    this.path = path;
    this.hashPartitionID = hashPartitionID;
    this.numHashPartitions = numHashPartitions;
    this.date = date;
  }

  /**
   * Loads all the information (tweet count, etc.) for this partition and day from HDFS.
   */
  public void load(FileSystem hdfs) throws IOException {
    FileStatus[] dailyBatchFiles = null;
    try {
      // listStatus() javadoc says it throws FileNotFoundException when path does not exist.
      // However, the actual implementations return null or an empty array instead.
      // We handle all 3 cases: null, empty array, or FileNotFoundException.
      dailyBatchFiles = hdfs.listStatus(path);
    } catch (FileNotFoundException e) {
      // don't do anything here and the day will be handled as empty.
    }

    if (dailyBatchFiles != null && dailyBatchFiles.length > 0) {
      for (FileStatus file : dailyBatchFiles) {
        String fileName = file.getPath().getName();
        if (fileName.equals(STATUS_COUNT_FILE_PREFIX)) {
          // zero tweets in this partition - this can happen for early days in 2006
          handleEmptyPartition();
        } else {
          Matcher matcher = STATUS_COUNT_FILE_PATTERN.matcher(fileName);
          if (matcher.matches()) {
            try {
              statusCount = Integer.parseInt(matcher.group(1));
              // Only adjustMinStatusId in production. For tests, this makes the tests harder to
              // understand.
              minStatusID = Config.environmentIsTest() ? Long.parseLong(matcher.group(2))
                  : adjustMinStatusId(Long.parseLong(matcher.group(2)), date);
              maxStatusID = Long.parseLong(matcher.group(3));
              hasStatusCountFile = true;
            } catch (NumberFormatException e) {
              // invalid file - ignore
              LOG.warn("Could not parse status count file name.", e);
            }
          }
        }
      }
    } else {
      // Partition folder does not exist. This case can happen for early days of twitter
      // where some partitions are empty. Set us to having a status count file, the validity of
      // the parent DailyStatusBatch will still be determined by whether there was a _SUCCESS file
      // in the day root.
      handleEmptyPartition();

      if (date.after(getEarliestDenseDay())) {
        LOG.error("Unexpected empty directory {} for {}", path, date);
      }
    }
  }

  private void handleEmptyPartition() {
    statusCount = 0;
    minStatusID = DailyStatusBatch.EMPTY_BATCH_STATUS_ID;
    maxStatusID = DailyStatusBatch.EMPTY_BATCH_STATUS_ID;
    hasStatusCountFile = true;
  }

  /**
   * Sometimes tweets are out-of-order (E.g. a tweet from Sep 2012 got into a
   * batch in July 2013). See SEARCH-1750 for more details.
   * This adjust the minStatusID if it is badly out-of-order.
   */
  @VisibleForTesting
  protected static long adjustMinStatusId(long minStatusID, Date date) {
    long dateTime = date.getTime();
    // If the daily batch is for a day before we started using snow flake IDs. Never adjust.
    if (!SnowflakeIdParser.isUsableSnowflakeTimestamp(dateTime)) {
      return minStatusID;
    }

    long earliestStartTime = dateTime - MAXIMUM_OUT_OF_ORDER_TOLERANCE_MILLIS;
    long minStatusTime = SnowflakeIdParser.getTimestampFromTweetId(minStatusID);
    if (minStatusTime < earliestStartTime) {
      long newMinId =  SnowflakeIdParser.generateValidStatusId(earliestStartTime, 0);
      LOG.info("Daily batch for " + date + " has badly out of order tweet: " + minStatusID
          + ". The minStatusID for the day this batch is adjusted to " + newMinId);
      return newMinId;
    } else {
      return minStatusID;
    }
  }

  /**
   * Returns a reader that reads tweets from the given directory.
   *
   * @param archiveSegment Determines the timeslice ID of all read tweets.
   * @param tweetsPath The path to the directory where the tweets for this day are stored.
   * @param documentFactory The ThriftIndexingEvent to TweetDocument converter.
   */
  public RecordReader<TweetDocument> getTweetReaders(
      ArchiveSegment archiveSegment,
      Path tweetsPath,
      DocumentFactory<ThriftIndexingEvent> documentFactory) throws IOException {
    RecordReader<TweetDocument> tweetDocumentReader =
        new TransformingRecordReader<>(
            createTweetReader(tweetsPath), new Function<ThriftIndexingEvent, TweetDocument>() {
          @Override
          public TweetDocument apply(ThriftIndexingEvent event) {
            return new TweetDocument(
                event.getSortId(),
                archiveSegment.getTimeSliceID(),
                EarlybirdThriftDocumentUtil.getCreatedAtMs(event.getDocument()),
                documentFactory.newDocument(event)
            );
          }
        });

    tweetDocumentReader.setExhaustStream(true);
    return tweetDocumentReader;
  }

  private RecordReader<ThriftIndexingEvent> createTweetReader(Path tweetsPath) throws IOException {
    if (date.before(START_DATE_INCLUSIVE)) {
      return new EmptyRecordReader<>();
    }

    List<RecordReader<ThriftIndexingEvent>> readers = Lists.newArrayList();
    FileSystem hdfs = HdfsUtil.getHdfsFileSystem();
    try {
      Path dayPath = new Path(tweetsPath, ArchiveHDFSUtils.dateToPath(date, "/"));
      Path partitionPath =
          new Path(dayPath, String.format("p_%d_of_%d", hashPartitionID, numHashPartitions));
      PathFilter pathFilter =
          Config.environmentIsTest() ? TXT_DATA_FILES_FILTER : LZO_DATA_FILES_FILTER;
      FileStatus[] files = hdfs.listStatus(partitionPath, pathFilter);
      for (FileStatus fileStatus : files) {
        String fileStatusPath = fileStatus.getPath().toString().replaceAll("file:/", "/");
        RecordReader<ThriftIndexingEvent> reader = createRecordReaderWithRetries(fileStatusPath);
        readers.add(reader);
      }
    } finally {
      IOUtils.closeQuietly(hdfs);
    }

    if (readers.isEmpty()) {
      return new EmptyRecordReader<>();
    }

    return new MergingSortedRecordReader<>(DESC_THRIFT_INDEXING_EVENT_COMPARATOR, readers);
  }

  private RecordReader<ThriftIndexingEvent> createRecordReaderWithRetries(String filePath)
      throws IOException {
    Predicate<ThriftIndexingEvent> recordFilter = getRecordFilter();
    int numTries = 0;
    while (true) {
      try {
        ++numTries;
        return new LzoThriftBlockFileReader<>(filePath, ThriftIndexingEvent.class, recordFilter);
      } catch (IOException e) {
        if (numTries < READER_INIT_IOEXCEPTION_RETRIES) {
          LOG.warn("Failed to open LzoThriftBlockFileReader for " + filePath + ". Will retry.", e);
        } else {
          LOG.error("Failed to open LzoThriftBlockFileReader for " + filePath
              + " after too many retries.", e);
          throw e;
        }
      }
    }
  }

  private Predicate<ThriftIndexingEvent> getRecordFilter() {
    return Config.environmentIsTest() ? null : input -> {
      if (input == null) {
        return false;
      }
      // We only guard against status IDs that are too small, because it is possible
      // for a very old tweet to get into today's batch, but not possible for a very
      // large ID (a future tweet ID that is not yet published) to get in today's
      // batch, unless tweet ID generation messed up.
      long statusId = input.getSortId();
      boolean keep = statusId >= minStatusID;
      if (!keep) {
        LOG.debug("Out of order documentId: {} minStatusID: {} Date: {} Path: {}",
            statusId, minStatusID, date, path);
        OUT_OF_ORDER_STATUSES_SKIPPED.increment();
      }
      return keep;
    };
  }

  /**
   * Returns the number of statuses in this batch
   */
  public int getStatusCount() {
    return statusCount;
  }

  /**
   * Was the _status_count file was found in this folder.
   */
  public boolean hasStatusCount() {
    return hasStatusCountFile;
  }

  public long getMinStatusID() {
    return minStatusID;
  }

  public long getMaxStatusID() {
    return maxStatusID;
  }

  public Date getDate() {
    return date;
  }

  public Path getPath() {
    return path;
  }

  /**
   * Check whether the partition is
   * . empty and
   * . it is disallowed (empty partition can only happen before 2010)
   * (Empty partition means that the directory is missing when scan happens.)
   *
   * @return true if the partition has no documents and it is not allowed.
   */
  public boolean isDisallowedEmptyPartition() {
    return hasStatusCountFile
        && statusCount == 0
        && minStatusID == DailyStatusBatch.EMPTY_BATCH_STATUS_ID
        && maxStatusID == DailyStatusBatch.EMPTY_BATCH_STATUS_ID
        && date.after(getEarliestDenseDay());
  }

  @Override
  public String toString() {
    return "PartitionedBatch[hashPartitionId=" + hashPartitionID
        + ",numHashPartitions=" + numHashPartitions
        + ",date=" + date
        + ",path=" + path
        + ",hasStatusCountFile=" + hasStatusCountFile
        + ",statusCount=" + statusCount + "]";
  }

  private Date getEarliestDenseDay() {
    return EarlybirdConfig.getDate("archive_search_earliest_dense_day");
  }
}