Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Flume >> mail # user >> Log Events get Lost - flume 1.3

Copy link to this message
RE: Log Events get Lost - flume 1.3
I am taking care of hot spotting since my rowkey format is appid:timestamp. Do we have any timeout facility with batch size as well, so that we don't have to wait for the batch to be filled in case of larger batch size if we are not have more frequency of log events at the agent.

Another query I do have, if I add some sequence number (or UUID) in rowkey, how I would be able to search in hbase by rowkeys?


From: Brock Noland [mailto:[EMAIL PROTECTED]]
Sent: Tuesday, April 16, 2013 3:18 PM
Subject: Re: Log Events get Lost - flume 1.3

As Israel said, you should not depend on timestamp being unique. I'd proceed with caution when using the sequence id approach if it will be the front of you key for the same reason timestamp at the front of your key can be a problem. More on that below.

You didn't specify where the timestamp was located in your key, but I'd like to warn you about an extremely common problem new users of HBase have. It's called "hot spotting" and using a timestamp as the first portion of the hey is nearly going to guarantee your tables "hot spot".  I'd suggest reading chapter 9 of HBase the definitive guide.


On Tue, Apr 16, 2013 at 2:02 PM, Israel Ekpo <[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>> wrote:

You can append or prefix a unique value to the nano time since you want each key to be unique per batch.

Here is the first approach:

private static AtomicLong idCounter = new AtomicLong();

public static String createSequenceID()
    return String.valueOf(idCounter.getAndIncrement());
You can also use the UUID random number generator to get the unique values

String uniqueID = UUID.randomUUID().toString();


I prefer the first option since it is more readable and helpful especially when you are debugging issues.

I hope this helps.

On 16 April 2013 14:36, Kumar, Deepak8 <[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>> wrote:
Hi Brock,
Thanks for assisting.

Actually we have an interceptor implementation through which we are generating our row key for hbase (hbase is sink). If we have larger batch size then the chances are that the timestamp is getting repeated in rowkey which would overwrite the rows in hbase.

Could you please guide me if we do have any work around so that I can have larger batchsize as well as the row key is not repeated. I am taking the count till nano timestamp.


  public Event intercept(Event event) {
//      eventCounter++;
    //env, logType, appId, logPath and logFileName
    Map<String, String> headers = event.getHeaders();
    long now = System.currentTimeMillis();
    String nowNano = Long.toString(System.nanoTime());
    //nowNano = nowNano.substring(nowNano.length()-5);

    headers.put(TIMESTAMP, Long.toString(now));
    headers.put(HOST_NAME, hostName);
    headers.put(ENV, env);
    headers.put(LOG_TYPE, logType);
    headers.put(APP_ID, appId);
    headers.put(LOG_FILE_PATH, logFilePath);
    headers.put(LOG_FILE_NAME, logFileName);
    headers.put(TIME_STAMP_NANO, nowNano);

    return event;

  public List<Event> intercept(List<Event> events) {
    for (Event event : events) {
    return events;
From: Brock Noland [mailto:[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>]
Sent: Tuesday, April 16, 2013 10:39 AM
Subject: Re: Log Events get Lost - flume 1.3


There are two issues with your configuration:

1) batch size of 1 with file channel is anti-pattern. This will result in extremely poor performance because the file channel will have to do an fsync() (expensive disk operation required to ensure no data loss) for each event. Your batch size should probably be in the hundreds or thousands.

2) tail -F *will* lose data. There is a writeup on this in documentation. If you care about your data, you will want to use Spooling Directory Source.

Issue #2 is being worsened by issue #1. Since you have such a low batch size, throughput of the file channel is extremely low. As tail -F results in no feedback to the tail process, more data than is being lost than would otherwise be the case due to the low channel throughput.

On Tue, Apr 16, 2013 at 3:16 AM, Kumar, Deepak8 <[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>> wrote:


I have 10 flume agents configured at a single machine. A single log file has frequency of 500 log events/sec. Hence in 10 log files the logs are getting generated as 5000 log events per second (5000/sec).

If my channel capacity is 1 million,  more than 70% of log events is lost! If I increase the channel capacity to 50 millions, then flume agent takes more than 24 hours to transfer the log events from source to sink.

The size of dataDir (agent.channels.fileChannel.dataDirs = /var/log/flume-ng/file-channel/data) is almost 2G all the time.

Could you please guide me the optimum configuration so that I don't miss any of log events & the transfer is also good enough. My flume-conf.properties has following contents:

agent.channels = fileChannel

agent.sinks = avroSink

# Each sink's type must be defined

agent.sinks.avroSink.type = avro

agent.sinks.avroSink.hostname = spnnq01.nam.nsroot.net<http://spnnq01.nam.nsroot.net>

agent.sinks.avroSink.port = 1442

agent.sinks.avroSink.batchSize = 1000

#Specify the channel the sink should use

agent.sinks.avroSink.channel = fileChannel

# Each channel's type is defined.

agent.channels.fileChannel.type = file

agent.channels.fileChannel.checkpointDir = /var/log/flume-ng/file-channel/checkpoint

agent.channels.fileChannel.dataDirs = /var/log/flume-ng/file-channel/data

agent.channels.fileChannel.transactionCapacity = 1000

agent.channels.fileChannel.checkpointInterval = 30000

agent.channels.fileChannel.maxFileSize = 2146435071<tel:2146435071>