Error Handlers

edit

Error Handlers are a beta feature in 6.2.0. This means that the APIs and configurations detailed here may be subject to breaking changes between releases while we improve the usability and clarity of the feature.

Elasticsearch for Apache Hadoop is designed to be a mostly hands off integration. Most of the features are managed through conventions and configurations and no substantial amount of code is required to get up and running with the connector. When it comes to exceptions, substantial effort has been put into handling the most common and expected errors from Elasticsearch. In the case where errors are unexpected or indicative of a real problem, the connector adopts a "fail-fast" approach. We realize that this approach is not the best for all users, especially those concerned with uptime of their jobs.

In these situations, users can handle unexpected errors by specifying the actions to take when encountering them. To this end, we have provided a set of APIs and extensions that users can implement in order to tailor the connector’s behavior toward encountering failures in the most common locations to their needs.

Error Handler Mechanics

edit

Each type of failure that can be handled in elasticsearch-hadoop has its own error handler API tailored to operation being performed. An error handler is simply a function that is called when an error occurs, and informs elasticsearch-hadoop how to proceed. Multiple error handlers may be specified for each type of failure. When a failure occurs, each error handler will be executed in the order that they are specified in the configuration.

An error handler is given information about the performed operation and any details about the error that was encountered. The handler may then acknowledge and consume the failure telling elasticsearch-hadoop to ignore the error. Alternatively, the handler may mark the error to be rethrown, potentially ending the job. Error handlers are also given the option to modify the operation’s parameters and retry it. Finally, the handler may also "pass" the failure on to the next handler in the list if it does not know how to handle it.

If every handler in the provided list of handlers chooses to "pass", it is marked as an unhandled error and the exceptions will be rethrown, potentially ending the job. The connector ships with a few default error handlers that take care of most use cases, but if you find that you need a more specific error handling strategy, you can always write your own.

In the following sections, we will detail the different types of error handlers, where they are called, how to configure them, and how to write your own if you need to.

Bulk Write Error Handlers

edit

When writing data, the connector batches up documents into a bulk request before sending them to Elasticsearch. In the response, Elasticsearch returns a status for each document sent, which may include rejections or failures. A common error encountered here is a rejection which means that the shard that the document was meant to be written to was too busy to accept the write. Other failures here might include documents that are refused because they do not conform to the current index mapping, or conflict with the current version of the document.

Elasticsearch for Apache Hadoop provides an API to handle document level errors from bulk responses. Error handlers for bulk writes are given:

  • The raw JSON bulk entry that was tried
  • Error message
  • HTTP status code for the document
  • Number of times that the current document has been sent to Elasticsearch

HTTP Retry Handler

edit

Always configured as the first error handler for bulk writes.

This handler checks the failures for common HTTP codes that can be retried. These codes are usually indicators that the shard the document would be written to is too busy to accept the write, and the document was rejected to shed load. This handler is always configured to be run first for bulk failures. All handlers that are configured by the user are placed in order after this one.

While the handler’s position in the list of error handlers cannot be modified, its behavior can be modified by adjusting the following configurations:

es.batch.write.retry.policy (default: simple)
Defines the policy for determining which http codes are able to be retried. The default value simple allows for 429 (Too Many Requests) and 503 (Unavailable) to be retried. Setting the value to none will allow no status codes to be retried.
es.batch.write.retry.count (default 3)
Number of retries for a given batch in case Elasticsearch is overloaded and data is rejected. Note that only the rejected data is retried. If there is still data rejected after the retries have been performed, the Hadoop job is cancelled (and fails). A negative value indicates infinite retries; be careful in setting this value as it can have unwanted side effects.
es.batch.write.retry.wait (default 10s)
Time to wait between batch write retries that are caused by bulk rejections.

Drop and Log Error Handler

edit

Default Handler Name: log

When this handler is invoked it logs a message containing the JSON bulk entry that failed, the error message, and any previous handler messages. After logging this message, the handler signals that the error has been acknowledged, thus consuming/ignoring it.

Available configurations for this handler:

es.write.rest.error.handler.log.logger.name (required)
The string name to use when creating the logger instance to log the errors. This setting is required if this handler is used.
es.write.rest.error.handler.log.logger.class (alternative to logger.name)
The class name to use when creating the logger instance to log the errors. This setting can be used instead of the required setting es.write.rest.error.handler.log.logger.name.
es.write.rest.error.handler.log.logger.level (default: WARN)
The logger level to use when logging the error message. Available options are FATAL, ERROR, WARN, INFO, DEBUG, and TRACE.

Abort on Failure Error Handler

edit

Default Handler Name: fail

When this handler is called it rethrows the error given to it and aborts. This handler is always loaded and automatically placed at the end of the list of error handlers.

There are no configurations for this handler.

Using Bulk Error Handlers

edit

To configure bulk error handlers, you must specify the handlers in order with the following properties.

Setting es.write.rest.error.handlers
Lists the names of the error handlers to use for bulk write error handling, and the order that they should be called on. Each default handler can be referenced by their handler name as the connector knows how to load them. Any handlers provided from users or third party code will need to have their handler names defined with the es.write.rest.error.handler. prefix.

For bulk write failures, the HTTP Retry built-in handler is always placed as the first error handler. Additionally, the Abort on Failure built-in handler is always placed as the last error handler to catch any unhandled errors. These two error handlers alone form the default bulk write error handling behavior for elasticsearch-hadoop, which matches the behavior from previous versions.

  1. HTTP Retry Built-In Handler: Retries benign bulk rejections and failures from Elasticsearch and passes any other error down the line
  2. Any configured user handlers will go here.
  3. Abort on Failure Built-In Handler: Rethrows the any errors it encounters

This behavior is modified by inserting handlers into the chain by using the handlers property. Let’s say that we want to log ALL errors and ignore them.

es.write.rest.error.handlers = log 

Specifying the default Drop and Log handler

With the above configuration, the handler list now looks like the following:

  1. HTTP Retry Handler
  2. Drop and Log Handler
  3. Abort on Failure Handler

As described above, the built-in log error handler has a required setting: What to use for the logger name. The logger used will respect whatever logging configuration you have in place, and thus needs a name for the logger to use:

es.write.rest.error.handlers = log 
es.write.rest.error.handler.log.logger.name = BulkErrors 

Specifying the default Drop and Log built-in handler

The Drop and Log built-in handler will log all errors to the BulkErrors logger

At this point, the Abort on Failure built-in handler is effectively ignored since the Drop and Log built-in handler will always mark an error as consumed. This practice can prove to be hazardous, as potentially important errors may simply be ignored. In many cases, it is preferable for users to write their own error handler to handle expected exceptions.

Writing Your Own Bulk Error Handlers

edit

Let’s say that you are streaming sensitive transaction data to Elasticsearch. In this scenario, your data is carefully versioned and you take advantage of Elasticsearch’s version system to keep from overwriting newer data with older data. Perhaps your data is distributed in a way that allows newer data to sneak in to Elasticsearch before some older bits of data. No worries, the version system will reject the older data and preserve the integrity of the data in Elasticsearch. The problem here is that your streaming job has failed because conflict errors were returned and the connector was unsure if you were expecting that.

Let’s write an error handler for this situation:

package org.myproject.myhandlers;

import org.elasticsearch.hadoop.handler.HandlerResult;
import org.elasticsearch.hadoop.rest.bulk.handler.BulkWriteErrorHandler;
import org.elasticsearch.hadoop.rest.bulk.handler.BulkWriteFailure;
import org.elasticsearch.hadoop.rest.bulk.handler.DelayableErrorCollector;

public class IgnoreConflictsHandler extends BulkWriteErrorHandler { 

    private static final Logger LOGGER = ...; 

    @Override
    public HandlerResult onError(BulkWriteFailure entry, DelayableErrorCollector<byte[]> collector) 
    throws Exception
    {
        if (entry.getResponseCode() == 409) { 
            LOGGER.warn("Encountered conflict response. Ignoring old data.");
            return HandlerResult.HANDLED; 
        }
        return collector.pass("Not a conflict response code."); 
    }
}

We create a class and extend the BulkWriteErrorHandler base class

Create a logger using preferred logging solution

Override the onError method which will be invoked with the error details

Check the response code from the error to see if it is 409 (Confict)

If it is a conflict, log the error and return HandlerResult.HANDLED to signal that the error is acknowledged

If the error is not a conflict we pass it along to the next error handler with the reason we couldn’t handle it

Before we can place this handler in the list of bulk write error handlers, we must register the handler class with a name in the settings using es.write.rest.error.handler.[HANDLER-NAME]:

Setting es.write.rest.error.handler.[HANDLER-NAME]
Create a new handler named HANDLER-NAME. The value of this property must be the binary name of the class to instantiate for this handler.

In this case, lets register a handler name for our ignore conflicts handler:

es.write.rest.error.handler.ignoreConflict = org.myproject.myhandlers.IgnoreConflictsHandler

Now that we have a name for the handler, we can use it in the handler list:

es.write.rest.error.handlers = ignoreConflict
es.write.rest.error.handler.ignoreConflict = org.myproject.myhandlers.IgnoreConflictsHandler

Now, your ignore conflict error handler will be invoked whenever a bulk failure occurs, and will instruct the connector that it is ok with ignoring conflict response codes from Elasticsearch.

Advanced Concepts

edit

What if instead of logging data and dropping it, what if you wanted to persist it somewhere for safe keeping? What if we wanted to pass properties into our handlers to parameterize their behavior? Lets create a handler that stores error information in a local file for later analysis.

package org.myproject.myhandlers;

import ...

import org.elasticsearch.hadoop.handler.HandlerResult;
import org.elasticsearch.hadoop.rest.bulk.handler.BulkWriteErrorHandler;
import org.elasticsearch.hadoop.rest.bulk.handler.BulkWriteFailure;
import org.elasticsearch.hadoop.rest.bulk.handler.DelayableErrorCollector;

public class OutputToFileHandler extends BulkWriteErrorHandler { 

    private OutputStream outputStream;   
    private BufferedWriter writer;

    @Override
    public void init(Properties properties) {   
        try {
            outputStream = new FileOutputStream(properties.getProperty("filename"));   
            writer = new BufferedWriter(new OutputStreamWriter(outputStream));
        } catch (FileNotFoundException e) {
            throw new RuntimeException("Could not open file", e);
        }
    }

    @Override
    public HandlerResult onError(BulkWriteFailure entry, DelayableErrorCollector<byte[]> collector)   
    throws Exception
    {
        writer.write("Code: " + entry.getResponseCode());
        writer.newLine();
        writer.write("Error: " + entry.getException().getMessage());
        writer.newLine();
        for (String message : entry.previousHandlerMessages()) {
            writer.write("Previous Handler: " + message);           
            writer.newLine();
        }
        writer.write("Attempts: " + entry.getNumberOfAttempts());
        writer.newLine();
        writer.write("Entry: ");
        writer.newLine();
        IOUtils.copy(entry.getEntryContents(), writer);
        writer.newLine();

        return HandlerResult.HANDLED; 
    }

    @Override
    public void close() {   
        try {
            writer.close();
            outputStream.close();
        } catch (IOException e) {
            throw new RuntimeException("Closing file failed", e);
        }
    }
}

Extend the BulkWriteErrorHandler base class

Some local state for writing data out to a file

We override the init method. Any properties for this handler are passed in here.

We are extracting the file to write to from the properties. We’ll see how to set this property below.

Overriding the onError method to define our behavior.

Write out the error information. This highlights all the available data provided by the BulkWriteFailure object.

Return the HANDLED result to signal that the error is handled.

Finally, close out any internally allocated resources.

Added to this handler are the init and close methods. The init method is called when the handler is first created at the start of the task and the close method is called when the task concludes. The init method accepts a properties parameter, which contains any handler specific properties set by using es.write.rest.error.handler.[HANDLER-NAME].[PROPERTY-NAME].

Setting es.write.rest.error.handler.[HANDLER-NAME].[PROPERTY-NAME]
Used to pass properties into handlers. HANDLER-NAME is the handler to be configured, and PROPERTY-NAME is the property to set for the handler.

In our use case, we will configure the our file logging error handler like so:

es.write.rest.error.handler.writeFile = org.myproject.myhandlers.OutputToFileHandler   
es.write.rest.error.handler.writeFile.filename = /path/to/some/output/file   

We register our new handler with the name writeFile

Now we set a property named filename for the writeFile handler. In the init method of the handler, this can be picked up by using filename as the property key.

Now to bring it all together with the previous example (ignoring conflicts):

es.write.rest.error.handlers = ignoreConflict,writeFile

es.write.rest.error.handler.ignoreConflict = org.myproject.myhandlers.IgnoreConflictsHandler

es.write.rest.error.handler.writeFile = org.myproject.myhandlers.OutputToFileHandler
es.write.rest.error.handler.writeFile.filename = /path/to/some/output/file

You now have a chain of handlers that retries bulk rejections by default (HTTP Retry built-in handler), then ignores any errors that are conflicts (our own ignore conflicts handler), then ignores any other errors by writing them out to a file (our own output to file handler).