Skip to main content
Loading

Batch Operations

A batch is a series of requests that are sent together to the database server. A batch groups multiple operations into one unit and passes it in a single network trip to each database node.

Batch advantages

  • Batches combine multiple record operations, including updates, deletes, reads and UDFs.
  • Batching allows any key-value operation, such as mixing record-level gets and deletes, as well as bin-level transaction operations including increment, prepend/append, Map and List operations, and bitwise operations.
  • When batching multiple updates, the number of connections between client and the server is reduced.
  • Batches optimize the use of network resources, such as, packets and network sockets.
  • Reduces Round Trip Time (RTT) when many primary keys are involved.
  • Backward compatible for applications using batch-read, batch-operate, and batch-exists.
  • Stores and retrieves multiple data points.

Batch use cases

  • Financial services (High performance computing)
  • Digital Media (Rendering and transcoding)
  • Internet of Things (Ingesting and analyzing IoT sensor data)

Batch workflow

The client groups the primary keys in the batch by a cluster node and creates a sub-batch request to each node. Batch requests occur in series or in parallel depending on the batch policy. Parallel requests in synchronous mode require extra threads, which are created or taken from a thread pool.

Batch requests use a single network socket to each cluster node, which helps with parallelizing requests. Multiple keys use one network request, which is beneficial for a large number of small records, but not as beneficial when the number of records per node is small, or the data returned per record is very large. Batch requests sometimes increase the latency of some requests, but only because clients normally wait until all the keys are retrieved from the cluster nodes before it returns control to the caller.

Some clients, such as the C client, deliver each record as soon as it arrives, allowing client applications to process data from the fastest server first. In the Java client, sync batch and async batch with RecordArrayListener wait until all responses arrive from the cluster nodes. Async batch calls the RecordSequenceListener to send back records one at a time as they are received.

Batch considerations

  • Batches are not multi record transactions.
  • There is no guarantee for the order of execution across nodes. The operations in a sub-batch can be processed in order, a behavior controlled by a batch policy flag. Otherwise, the operations in a sub-batch execute in parallel on the node.
  • There is no rollback processing for failed operations in a batch.
  • You can configure the batch policy to stop or continue batch processing if an operation fails.
  • Use the operate command to combine multiple changes with the same key in a single batch transaction.
  • Starting with Aeroskipe 6.0, you can combine any type of read and write operations against one or more keys.
  • Batch transactions will use more resources than single record transactions end-to-end (client and server) unless batches are really large. It is recommended to use single record transactions if the batch size is smaller.

Inlining batches

If inline is set to true through flags in the batch policy, the node executes every operation in its sub-batch, one after the other.

  • Batch Policy allowInline controls whether to inline sub-batch operations where the keys are in an in-memory namespace. The default value is true.
  • Batch Policy allowInlineSSD controls whether to inline sub-batch operations where the keys are in an SSD based namespace. The default value is false.
  • For batch transactions with smaller records, for example, 1KiB per record, using inline processing is faster for in-memory namespaces.
  • If it is not inline, the sub-batch operations are split and executed by different threads.
  • Inlining sub-batches does not tend to improve latency when the keys are in an SSD-based namespace. You should benchmark to compare performance.
  • When a sub-batch is inlined, one thread executes the operations. The thread is not released until the sub-batch processing is complete. Large inlined batches may divert server resources toward batch operations over single-record operations.

Filtering batch operations

You can attach a filter expression to any batch operation. The server applies the filter to each record in the batch to determine whether the operation should proceed.

Supported batch requests

Batch reads

  • Batch get: Reads multiple records, or optionally a selection of bins from those records.
  • Batch exists: Verifies metadata whether the specified keys exist.
  • Batch getHeader: Reads record metadata (expiration/TTL and generation) only. It does not return bin data.
  • Batch operate: Executes a transaction of read operations against multiple records.

Batch writes

  • Aerospike release 6.0 introduces batch writes in addition to batch reads. This includes updates, deletes, UDFs, and multi-operation transactions (operate) without limits on write operations.
  • Batch writes allow write operations against any keys.
  • Batch writes can process large numbers of updates in a single operation using less connections.

Multiple Batch Operations to the Same Key in a Batch

  • Unless you inline the batch request to be serviced by the same service thread, multiple batch ops to same key, e.g. [K1, K3...K1, K5....K1, K2] will get distributed to different service threads.
  • Client library or the server cannot decipher your intent of operations on K1 and consolidate them into one batch transaction on K1.
  • Each operation on K1 can proceed, potentially, out of order when distributed to different service threads.

If you don't want this situation, either consolidate operations for the same key into one key entry in the batch, or inline the batch request.  

Code examples of batch operations

The following examples use a Java client to perform read, write, delete, exist, and mixed read/write operations on a batch. The syntax differs per Aerospike client, based on the needs of the operating system.

Example: Batch read records

/**
* Read records in one batch.
*/
private void batchReads (
AerospikeClient client,
Parameters params,
String keyPrefix,
String binName,
int size
) throws Exception {
// Batch gets into one call.
Key[] keys = new Key[size];
for (int i = 0; i < size; i++) {
keys[i] = new Key(params.namespace, params.set, keyPrefix + (i + 1));
}

Record[] records = client.get(null, keys, binName);

for (int i = 0; i < records.length; i++) {
Key key = keys[i];
Record record = records[i];
Level level = Level.ERROR;
Object value = null;

if (record != null) {
level = Level.INFO;
value = record.getValue(binName);
}
console.write(level, "Record: ns=%s set=%s key=%s bin=%s value=%s",
key.namespace, key.setName, key.userKey, binName, value);
}

if (records.length != size) {
console.error("Record size mismatch. Expected %d. Received %d.", size, records.length);
}
}

Example: Batch read operation

private void batchListOperate(AerospikeClient client, Parameters params) {
console.info("batchListOperate");
Key[] keys = new Key[RecordCount];
for (int i = 0; i < RecordCount; i++) {
keys[i] = new Key(params.namespace, params.set, KeyPrefix + (i + 1));
}

// Get size and last element of list bin for all records.
Record[] records = client.get(null, keys,
ListOperation.size(BinName3),
ListOperation.getByIndex(BinName3, -1, ListReturnType.VALUE)
);

for (int i = 0; i < records.length; i++) {
Record record = records[i];
//System.out.println(record);

List<?> results = record.getList(BinName3);
long size = (Long)results.get(0);
Object val = results.get(1);

console.info("Result[%d]: %d,%s", i, size, val);
}
}

Example: Batch read/write operations

/*
* Perform list read/write operations in one batch.
*/
private void batchListWriteOperate(AerospikeClient client, Parameters params) {
console.info("batchListWriteOperate");
Key[] keys = new Key[RecordCount];
for (int i = 0; i < RecordCount; i++) {
keys[i] = new Key(params.namespace, params.set, KeyPrefix + (i + 1));
}

// Add integer to list and get size and last element of list bin for all records.
BatchResults bresults = client.operate(null, null, keys,
ListOperation.append(ListPolicy.Default, BinName3, Value.get(999)),
ListOperation.size(BinName3),
ListOperation.getByIndex(BinName3, -1, ListReturnType.VALUE)
);

for (int i = 0; i < bresults.records.length; i++) {
BatchRecord br = bresults.records[i];
Record rec = br.record;

if (rec != null) {
List<?> results = rec.getList(BinName3);
long size = (Long)results.get(1);
Object val = results.get(2);

console.info("Result[%d]: %d,%s", i, size, val);
}
else {
console.info("Result[%d]: error: %s", i, ResultCode.getResultString(br.resultCode));
}
}
}

Example: Batch exist operation

/*
* Check existence of records in one batch.
*/
private void batchExists (
AerospikeClient client,
Parameters params,
String keyPrefix,
int size
) throws Exception {
// Batch into one call.
Key[] keys = new Key[size];
for (int i = 0; i < size; i++) {
keys[i] = new Key(params.namespace, params.set, keyPrefix + (i + 1));
}

boolean[] existsArray = client.exists(null, keys);

for (int i = 0; i < existsArray.length; i++) {
Key key = keys[i];
boolean exists = existsArray[i];
console.info("Record: ns=%s set=%s key=%s exists=%s",
key.namespace, key.setName, key.userKey, exists);
}
}

Example: Batch delete operation

public void batchDelete() {
// Define keys
Key[] keys = new Key[] {
new Key(args.namespace, args.set, 10000),
new Key(args.namespace, args.set, 10001)
};

// Ensure keys exists
boolean[] exists = client.exists(null, keys);
assertTrue(exists[0]);
assertTrue(exists[1]);

// Delete keys
BatchResults br = client.delete(null, null, keys);
assertTrue(br.status);

// Ensure keys do not exist
exists = client.exists(null, keys);
assertFalse(exists[0]);
assertFalse(exists[1]);
}

Example: Batch read/write mixed operation

/*
* Read/Write records using varying operations in one batch.
*/
private void batchWriteOperateComplex(AerospikeClient client, Parameters params) {
console.info("batchWriteOperateComplex");
Expression wexp1 = Exp.build(Exp.add(Exp.intBin(BinName1), Exp.intBin(BinName2), Exp.val(1000)));
Expression rexp1 = Exp.build(Exp.mul(Exp.intBin(BinName1), Exp.intBin(BinName2)));
Expression rexp2 = Exp.build(Exp.add(Exp.intBin(BinName1), Exp.intBin(BinName2)));
Expression rexp3 = Exp.build(Exp.sub(Exp.intBin(BinName1), Exp.intBin(BinName2)));

Operation[] ops1 = Operation.array(
Operation.put(new Bin(BinName4, 100)),
ExpOperation.read(ResultName1, rexp1, ExpReadFlags.DEFAULT));

Operation[] ops2 = Operation.array(ExpOperation.read(ResultName1, rexp1, ExpReadFlags.DEFAULT));
Operation[] ops3 = Operation.array(ExpOperation.read(ResultName1, rexp2, ExpReadFlags.DEFAULT));

Operation[] ops4 = Operation.array(
ExpOperation.write(BinName1, wexp1, ExpWriteFlags.DEFAULT),
ExpOperation.read(ResultName1, rexp3, ExpReadFlags.DEFAULT));

Operation[] ops5 = Operation.array(
ExpOperation.read(ResultName1, rexp2, ExpReadFlags.DEFAULT),
ExpOperation.read(ResultName2, rexp3, ExpReadFlags.DEFAULT));

List<BatchRecord> records = new ArrayList<BatchRecord>();
records.add(new BatchWrite(new Key(params.namespace, params.set, KeyPrefix + 1), ops1));
records.add(new BatchRead(new Key(params.namespace, params.set, KeyPrefix + 2), ops2));
records.add(new BatchRead(new Key(params.namespace, params.set, KeyPrefix + 3), ops3));
records.add(new BatchWrite(new Key(params.namespace, params.set, KeyPrefix + 4), ops4));
records.add(new BatchRead(new Key(params.namespace, params.set, KeyPrefix + 5), ops5));
records.add(new BatchDelete(new Key(params.namespace, params.set, KeyPrefix + 6)));

// Execute batch.
client.operate(null, records);

// Show results.
int i = 0;
for (BatchRecord record : records) {
Record rec = record.record;

if (rec != null) {
Object v1 = rec.getValue(ResultName1);
Object v2 = rec.getValue(ResultName2);
console.info("Result[%d]: %s, %s", i, v1, v2);
}
else {
console.info("Result[%d]: error: %s", i, ResultCode.getResultString(record.resultCode));
}
i++;
}
}

Log examples

A batch information coming from batch-sub. Stats on operations including batches for the namespace called a test.

Ticker log line changes:

\\ Example:

{test} batch-sub: tsvc (0,0) proxy (0,0,0) read (959,0,0,51,1) write (0,0,0,0) delete (0,0,0,0,0) udf (0,0,0,0) lang (0,0,0,0)

When the cluster size changes you might also see proxy events included in a batch.

\\Example:
{test} from-proxy-batch-sub: tsvc (0,0) read (959,0,0,51,1) write (0,0,0,0) delete (0,0,0,0,0) udf (0,0,0,0) lang (0,0,0,0)

Batch specific errors

ValueErrorDescription
150AS_ERR_BATCH_DISABLEDBatch functionality has been disabled by configuring batch-index-threads to 0
152AS_ERR_BATCH_QUEUES_FULLAll batch queues are full. Controlled by the batch-max-buffers-per-queue configuration parameter

Refer to Error Codes.

Known Issues or Limitations

Batch writes were not supported prior to the server 6.0 version.

Required Client Versions

For full compatibility the following clients are required:

  • Java client 6.0.0 or later
  • Go client 6.0.0 or later
  • C client 6.0.0 or later
  • C# client 5.0.0 or later

Client References

Refer to these topics for language-specific code examples: