vlambda博客
学习文章列表

private HoodieRecordLocation getNewRecordLocation(String partitionPath) { // // 通过 recordKey 得到 partitionPath，获取对应分区的小文件信息，观察下一个代码块 final BucketInfo bucketInfo = this.bucketAssigner.addInsert(partitionPath); final HoodieRecordLocation location; switch (bucketInfo.getBucketType()) { case INSERT: // This is an insert bucket, use HoodieRecordLocation instant time as "I". // Downstream operators can then check the instant time to know whether // a record belongs to an insert bucket. location = new HoodieRecordLocation("I", bucketInfo.getFileIdPrefix()); break; case UPDATE: location = new HoodieRecordLocation("U", bucketInfo.getFileIdPrefix()); break; default: throw new AssertionError(); } return location; }

public BucketInfo addInsert(String partitionPath) { // 获取小文件 SmallFileAssign smallFileAssign = getSmallFileAssign(partitionPath);
 // assign 判断小文件是否还能再分配，不能超过 totalAssgin if (smallFileAssign != null && smallFileAssign.assign()) { return new BucketInfo(BucketType.UPDATE, smallFileAssign.getFileId(), partitionPath); }
 // 下面就是创建新的 fileId 写入 ... }


private synchronized SmallFileAssign getSmallFileAssign(String partitionPath) { // 判断是否缓存了对应分区路径的小文件信息 if (smallFileAssignMap.containsKey(partitionPath)) { return smallFileAssignMap.get(partitionPath); } // writeProfile.getSmallFiles 获取小文件，观察下一个代码块，然后回来 List<SmallFile> smallFiles = smallFilesOfThisTask(writeProfile.getSmallFiles(partitionPath)); if (smallFiles.size() > 0) { LOG.info("For partitionPath : " + partitionPath + " Small Files => " + smallFiles); // 重点关注： //  //注意这里：小文件返回后构造 assignState, 在初始化 SmallFileAssignState 时，会通过计算历史的平均每行数据的大小，如果没有，默认 1024 byte 作为每一行大小。 // 然后 (文件最大配置 - 小文件大小) / 平均行大小 = 这个小文件还能分配的行数（totalUnassigned // 后续的 processElement 每次就会调用 SmallFileAssignState assigned 分配（如果 partitionUrl 相同），直到分配完 // // // 重点关注： SmallFileAssignState[] states = smallFiles.stream() .map(smallFile -> new SmallFileAssignState(config.getParquetMaxFileSize(), smallFile, writeProfile.getAvgSize())) .toArray(SmallFileAssignState[]::new); SmallFileAssign assign = new SmallFileAssign(states); smallFileAssignMap.put(partitionPath, assign); return assign; } smallFileAssignMap.put(partitionPath, null); return null; }

public synchronized List<SmallFile> getSmallFiles(String partitionPath) { // lookup the cache first if (smallFilesMap.containsKey(partitionPath)) { return smallFilesMap.get(partitionPath); }
 List<SmallFile> smallFiles = new ArrayList<>(); if (config.getParquetSmallFileLimit() <= 0) { this.smallFilesMap.put(partitionPath, smallFiles); return smallFiles; }
 // 获取小文件，调用 smallFilesProfile smallFiles = smallFilesProfile(partitionPath); this.smallFilesMap.put(partitionPath, smallFiles); return smallFiles; }

// 非 MOR 表实现，MOR 表调用 DeltaWriteProfile.smallFIleProfileprotected List<SmallFile> smallFilesProfile(String partitionPath) { // smallFiles only for partitionPath List<SmallFile> smallFileLocations = new ArrayList<>();
 HoodieTimeline commitTimeline = metaClient.getCommitsTimeline().filterCompletedInstants();
 if (!commitTimeline.empty()) { // if we have some commits HoodieInstant latestCommitTime = commitTimeline.lastInstant().get(); // 获取指定分区下的所有文件（使用 Metadata 获取的 fsView） List<HoodieBaseFile> allFiles = fsView .getLatestBaseFilesBeforeOrOn(partitionPath, latestCommitTime.getTimestamp()).collect(Collectors.toList());
 for (HoodieBaseFile file : allFiles) { // 过滤出满足条件的文件 // 小于 hoodie.parquet.small.file.limit 默认 100M，并且大于 0 的文件 if (file.getFileSize() < config.getParquetSmallFileLimit() && file.getFileSize() > 0) { String filename = file.getFileName(); SmallFile sf = new SmallFile(); sf.location = new HoodieRecordLocation(FSUtils.getCommitTime(filename), FSUtils.getFileId(filename)); sf.sizeBytes = file.getFileSize(); smallFileLocations.add(sf); } } }
 return smallFileLocations; }

@Override protected List<SmallFile> smallFilesProfile(String partitionPath) { // smallFiles only for partitionPath List<SmallFile> smallFileLocations = new ArrayList<>();
 // Init here since this class (and member variables) might not have been initialized HoodieTimeline commitTimeline = metaClient.getCommitsTimeline().filterCompletedInstants();
 // Find out all eligible small file slices if (!commitTimeline.empty()) { HoodieInstant latestCommitTime = commitTimeline.lastInstant().get(); // find the smallest file in partition and append to it List<FileSlice> allSmallFileSlices = new ArrayList<>(); // If we can index log files, we can add more inserts to log files for fileIds including those under // pending compaction. // 获取 base_file + log_file 的文件偏 List<FileSlice> allFileSlices = fsView.getLatestFileSlicesBeforeOrOn(partitionPath, latestCommitTime.getTimestamp(), true) .collect(Collectors.toList()); for (FileSlice fileSlice : allFileSlices) { // 判断是否满足小文件的条件  // (baseFileSize + totalLogFileSize * ratio) < hoodie.parquet.max.file.size(120M) // 这里的 ratio 是 hoodie.logfile.to.parquet.compression.ratio，默认 0.35 if (isSmallFile(fileSlice)) { allSmallFileSlices.add(fileSlice); } } // Create SmallFiles from the eligible file slices for (FileSlice smallFileSlice : allSmallFileSlices) { SmallFile sf = new SmallFile(); if (smallFileSlice.getBaseFile().isPresent()) { // TODO : Move logic of file name, file id, base commit time handling inside file slice String filename = smallFileSlice.getBaseFile().get().getFileName(); sf.location = new HoodieRecordLocation(FSUtils.getCommitTime(filename), FSUtils.getFileId(filename)); sf.sizeBytes = getTotalFileSize(smallFileSlice); smallFileLocations.add(sf); } else { smallFileSlice.getLogFiles().findFirst().ifPresent(logFile -> { // in case there is something error, and the file slice has no log file sf.location = new HoodieRecordLocation(FSUtils.getBaseCommitTimeFromLogPath(logFile.getPath()), FSUtils.getFileIdFromLogPath(logFile.getPath())); sf.sizeBytes = getTotalFileSize(smallFileSlice); smallFileLocations.add(sf); }); } } } return smallFileLocations; }