Best Practices

Write records using Fat Client or Thin Client

Fat Client is the writer in distributedlog core library which talks to ZooKeeper and BookKeeper directly, while Thin Client is the write proxy client which talks to write proxy service.

It is strongly recommended that writing records via Write Proxy service rather than using core library directly. Because using Thin Client has following benefits:

  • Thin Client is purely thrift RPC based client. It doesn't talk to zookeeper and bookkeeper directly and less complicated.
  • Write Proxy manages ownerships of log writers. Thin Client doesn't have to deal with ownerships of log writers.
  • Thin Client is more upgrade-friendly than Fat Client.

The only exception to use distributedlog core library directly is when you application requires:

  • Write Ordering. Write Ordering means all the writes issued by the writer should be written in a strict order in the log. Write Proxy service could only guarantee Read Ordering. Read Ordering means the write proxies will write the write requests in their receiving order and gurantee the data seen by all the readers in same order.
  • Ownership Management. If the application already has any kind of ownership management, like master-slave, it makes more sense that it uses distributedlog core library directly.

How to position reader by time

Sometimes, application wants to read data by time, like read data from 2 hours ago. This could be done by positioning the reader using Transaction ID, if the Transaction ID is the timestamp (All the streams produced by Write Proxy use timestamp as Transaction ID).

DistributedLogManager dlm = ...;

long timestamp = System.currentTimeMillis();
long startTxId = timestamp - TimeUnit.MILLISECONDS.convert(2, TimeUnit.HOURS);
AsyncLogReader reader = Await.result(dlm.openAsyncLogReader(startTxId));

How to seal a stream

Typically, DistributedLog is used as endless streams. In some use cases, application wants to seal the stream. So writers can't write more data into the log stream and readers could receive notifications about the stream has been sealed.

Write could seal a log stream as below:

DistributedLogManager dlm = ...;

LogWriter writer = dlm.startLogSegmentNonPartitioned;
// writer writes bunch of records

// writer seals the stream

Reader could detect a stream has been sealed as below:

DistributedLogManager dlm = ...;

LogReader reader = dlm.getInputStream(1L);
LogRecord record;
try {
    while ((record = reader.readNext(false)) != null) {
        // process the record
} catch (EndOfStreamException eos) {
    // the stream has been sealed