Examples with AggregateAndSortData uk.gov.gchq.gaffer.parquetstore.operation.handler.utilities.AggregateAndSortData used on opensource projects

Example 1 with AggregateAndSortData

use of uk.gov.gchq.gaffer.parquetstore.operation.handler.utilities.AggregateAndSortData in project Gaffer by gchq.

the class AddElementsHandler method addElements.

private void addElements(final AddElements addElementsOperation, final Context context, final ParquetStore store) throws OperationException {
    // Set up
    final FileSystem fs = store.getFS();
    final Schema schema = store.getSchema();
    final SchemaUtils schemaUtils = store.getSchemaUtils();
    final SparkSession spark = SparkContextUtil.getSparkSession(context, store.getProperties());
    final ExecutorService threadPool = createThreadPool(spark, store.getProperties());
    final GraphPartitioner currentGraphPartitioner = store.getGraphPartitioner();
    SparkParquetUtils.configureSparkForAddElements(spark, store.getProperties());
    // Write data from addElementsOperation split by group and partition (NB this uses the existing partitioner -
    // adding elements using this operation does not effect the partitions).
    final String tmpDirectory = store.getTempFilesDir();
    final BiFunction<String, Integer, String> directoryForGroupAndPartitionId = (group, partitionId) -> tmpDirectory + "/unsorted_unaggregated_new" + "/group=" + group + "/partition=" + partitionId;
    final BiFunction<String, Integer, String> directoryForGroupAndPartitionIdForReversedEdges = (group, partitionId) -> tmpDirectory + "/unsorted_unaggregated_new" + "/reversed-group=" + group + "/partition=" + partitionId;
    LOGGER.info("Calling WriteUnsortedData to add elements");
    LOGGER.trace("currentGraphPartitioner is {}", currentGraphPartitioner);
    new WriteUnsortedData(store, currentGraphPartitioner, directoryForGroupAndPartitionId, directoryForGroupAndPartitionIdForReversedEdges).writeElements(addElementsOperation.getInput());
    // For every group and partition, aggregate the new data with the old data and then sort
    final BiFunction<String, Integer, String> directoryForSortedResultsForGroupAndPartitionId = (group, partitionId) -> tmpDirectory + "/sorted_new_old_merged" + "/group=" + group + "/partition=" + partitionId;
    final BiFunction<String, Integer, String> directoryForSortedResultsForGroupAndPartitionIdForReversedEdges = (group, partitionId) -> tmpDirectory + "/sorted_new_old_merged" + "/REVERSED-group=" + group + "/partition=" + partitionId;
    final List<Callable<CallableResult>> tasks = new ArrayList<>();
    for (final String group : schema.getGroups()) {
        final List<Partition> partitions = currentGraphPartitioner.getGroupPartitioner(group).getPartitions();
        for (final Partition partition : partitions) {
            final List<String> inputFiles = new ArrayList<>();
            // New data
            inputFiles.add(directoryForGroupAndPartitionId.apply(group, partition.getPartitionId()));
            // Old data
            inputFiles.add(store.getFile(group, partition));
            final String outputDir = directoryForSortedResultsForGroupAndPartitionId.apply(group, partition.getPartitionId());
            final AggregateAndSortData task = new AggregateAndSortData(schemaUtils, fs, inputFiles, outputDir, group, group + "-" + partition.getPartitionId(), false, store.getProperties().getCompressionCodecName(), spark);
            tasks.add(task);
            LOGGER.info("Created AggregateAndSortData task for group {}, partition {}", group, partition.getPartitionId());
        }
    }
    for (final String group : schema.getEdgeGroups()) {
        final List<Partition> partitions = currentGraphPartitioner.getGroupPartitionerForReversedEdges(group).getPartitions();
        for (final Partition partition : partitions) {
            final List<String> inputFiles = new ArrayList<>();
            // New data
            inputFiles.add(directoryForGroupAndPartitionIdForReversedEdges.apply(group, partition.getPartitionId()));
            // Old data
            inputFiles.add(store.getFileForReversedEdges(group, partition));
            final String outputDir = directoryForSortedResultsForGroupAndPartitionIdForReversedEdges.apply(group, partition.getPartitionId());
            final AggregateAndSortData task = new AggregateAndSortData(schemaUtils, fs, inputFiles, outputDir, group, "reversed-" + group + "-" + partition.getPartitionId(), true, store.getProperties().getCompressionCodecName(), spark);
            tasks.add(task);
            LOGGER.info("Created AggregateAndSortData task for reversed edge group {}, partition {}", group, partition.getPartitionId());
        }
    }
    try {
        LOGGER.info("Invoking {} AggregateAndSortData tasks", tasks.size());
        final List<Future<CallableResult>> futures = threadPool.invokeAll(tasks);
        for (final Future<CallableResult> future : futures) {
            final CallableResult result = future.get();
            LOGGER.info("Result {} from task", result);
        }
    } catch (final InterruptedException e) {
        throw new OperationException("InterruptedException running AggregateAndSortData tasks", e);
    } catch (final ExecutionException e) {
        throw new OperationException("ExecutionException running AggregateAndSortData tasks", e);
    }
    try {
        // Move results to a new snapshot directory (the -tmp at the end allows us to add data to the directory,
        // and then when this is all finished we rename the directory to remove the -tmp; this allows us to make
        // the replacement of the old data with the new data an atomic operation and ensures that a get operation
        // against the store will not read the directory when only some of the data has been moved there).
        final long snapshot = System.currentTimeMillis();
        final String newDataDir = store.getDataDir() + "/" + ParquetStore.getSnapshotPath(snapshot) + "-tmp";
        LOGGER.info("Moving aggregated and sorted data to new snapshot directory {}", newDataDir);
        fs.mkdirs(new Path(newDataDir));
        for (final String group : schema.getGroups()) {
            final Path groupDir = new Path(newDataDir, ParquetStore.getGroupSubDir(group, false));
            fs.mkdirs(groupDir);
            LOGGER.info("Created directory {}", groupDir);
        }
        for (final String group : schema.getEdgeGroups()) {
            final Path groupDir = new Path(newDataDir, ParquetStore.getGroupSubDir(group, true));
            fs.mkdirs(groupDir);
            LOGGER.info("Created directory {}", groupDir);
        }
        for (final String group : schema.getGroups()) {
            final String groupDir = newDataDir + "/" + ParquetStore.getGroupSubDir(group, false);
            final List<Partition> partitions = currentGraphPartitioner.getGroupPartitioner(group).getPartitions();
            for (final Partition partition : partitions) {
                final Path outputDir = new Path(directoryForSortedResultsForGroupAndPartitionId.apply(group, partition.getPartitionId()));
                if (!fs.exists(outputDir)) {
                    LOGGER.info("Not moving data for group {}, partition id {} as the outputDir {} does not exist", group, partition.getPartitionId(), outputDir);
                } else {
                    // One .parquet file and one .parquet.crc file
                    final FileStatus[] status = fs.listStatus(outputDir, path -> path.getName().endsWith(".parquet"));
                    if (1 != status.length) {
                        LOGGER.error("Didn't find one Parquet file in path {} (found {} files)", outputDir, status.length);
                        throw new OperationException("Expected to find one Parquet file in path " + outputDir + " (found " + status.length + " files)");
                    } else {
                        final Path destination = new Path(groupDir, ParquetStore.getFile(partition.getPartitionId()));
                        LOGGER.info("Renaming {} to {}", status[0].getPath(), destination);
                        fs.rename(status[0].getPath(), destination);
                    }
                }
            }
        }
        for (final String group : schema.getEdgeGroups()) {
            final String groupDir = newDataDir + "/" + ParquetStore.getGroupSubDir(group, true);
            final List<Partition> partitions = currentGraphPartitioner.getGroupPartitionerForReversedEdges(group).getPartitions();
            for (final Partition partition : partitions) {
                final Path outputDir = new Path(directoryForSortedResultsForGroupAndPartitionIdForReversedEdges.apply(group, partition.getPartitionId()));
                if (!fs.exists(outputDir)) {
                    LOGGER.info("Not moving data for reversed edge group {}, partition id {} as the outputDir {} does not exist", group, partition.getPartitionId(), outputDir);
                } else {
                    // One .parquet file and one .parquet.crc file
                    final FileStatus[] status = fs.listStatus(outputDir, path -> path.getName().endsWith(".parquet"));
                    if (1 != status.length) {
                        LOGGER.error("Didn't find one Parquet file in path {} (found {} files)", outputDir, status.length);
                        throw new OperationException("Expected to find one Parquet file in path " + outputDir + " (found " + status.length + " files)");
                    } else {
                        final Path destination = new Path(groupDir, ParquetStore.getFile(partition.getPartitionId()));
                        LOGGER.info("Renaming {} to {}", status[0].getPath(), destination);
                        fs.rename(status[0].getPath(), destination);
                    }
                }
            }
        }
        // Delete temporary data directory
        LOGGER.info("Deleting temporary directory {}", tmpDirectory);
        fs.delete(new Path(tmpDirectory), true);
        // Write out graph partitioner (unchanged from previous one)
        final Path newGraphPartitionerPath = new Path(newDataDir + "/graphPartitioner");
        final FSDataOutputStream stream = fs.create(newGraphPartitionerPath);
        LOGGER.info("Writing graph partitioner to {}", newGraphPartitionerPath);
        new GraphPartitionerSerialiser().write(currentGraphPartitioner, stream);
        stream.close();
        // Move snapshot-tmp directory to snapshot
        final String directoryWithoutTmp = newDataDir.substring(0, newDataDir.lastIndexOf("-tmp"));
        LOGGER.info("Renaming {} to {}", newDataDir, directoryWithoutTmp);
        fs.rename(new Path(newDataDir), new Path(directoryWithoutTmp));
        // Set snapshot on store to new value
        LOGGER.info("Updating latest snapshot on store to {}", snapshot);
        store.setLatestSnapshot(snapshot);
    } catch (final IOException | StoreException e) {
        throw new OperationException("IOException moving results files into new snapshot directory", e);
    }
}

Example 2 with AggregateAndSortData

use of uk.gov.gchq.gaffer.parquetstore.operation.handler.utilities.AggregateAndSortData in project Gaffer by gchq.

the class AggregateAndSortDataTest method test.

@Test
public void test(@TempDir java.nio.file.Path tempDir) throws Exception {
    // Given
    final FileSystem fs = FileSystem.get(new Configuration());
    final SchemaUtils schemaUtils = new SchemaUtils(TestUtils.gafferSchema("schemaUsingLongVertexType"));
    final String file1 = tempDir.resolve("inputdata1.parquet").toString();
    final String file2 = tempDir.resolve("inputdata2.parquet").toString();
    writeData(file1, schemaUtils);
    writeData(file2, schemaUtils);
    final SparkSession sparkSession = SparkSessionProvider.getSparkSession();
    final List<String> inputFiles = new ArrayList<>(Sets.newHashSet(file1, file2));
    final String outputFolder = tempDir.resolve("aggregated").toString();
    // When
    new AggregateAndSortData(schemaUtils, fs, inputFiles, outputFolder, TestGroups.ENTITY, "test", false, CompressionCodecName.GZIP, sparkSession).call();
    // Then
    assertTrue(fs.exists(new Path(outputFolder)));
    final Row[] results = (Row[]) sparkSession.read().parquet(outputFolder).collect();
    // Should be sorted by vertex and date
    for (int i = 0; i < 40; i++) {
        assertEquals((long) i / 2, (long) results[i].getAs(ParquetStore.VERTEX));
        assertEquals(i % 2 == 0 ? 'b' : 'a', ((byte[]) results[i].getAs("byte"))[0]);
        assertEquals(i % 2 == 0 ? 8f : 6f, results[i].getAs("float"), 0.01f);
        assertEquals(11L * 2 * (i / 2), (long) results[i].getAs("long"));
        assertEquals(i % 2 == 0 ? 14 : 12, (int) results[i].getAs("short"));
        assertEquals(i % 2 == 0 ? 100000L : 200000L, (long) results[i].getAs("date"));
        assertEquals(2, (int) results[i].getAs("count"));
        assertArrayEquals(i % 2 == 0 ? new String[] { "A", "C" } : new String[] { "A", "B" }, (String[]) ((WrappedArray<String>) results[i].getAs("treeSet")).array());
        final FreqMap mergedFreqMap1 = new FreqMap();
        mergedFreqMap1.put("A", 2L);
        mergedFreqMap1.put("B", 2L);
        final FreqMap mergedFreqMap2 = new FreqMap();
        mergedFreqMap2.put("A", 2L);
        mergedFreqMap2.put("C", 2L);
        assertEquals(JavaConversions$.MODULE$.mapAsScalaMap(i % 2 == 0 ? mergedFreqMap2 : mergedFreqMap1), results[i].getAs("freqMap"));
    }
}