Error Handlers

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:

There are a few default error handlers provided by the connector:

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:

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:

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.

Default Handler Name: es

When this handler is invoked it converts the error given into a JSON document and inserts it into an Elasticsearch index. When this indexing operation is complete, the handler can be configured to signal that the error is handled (default), or pass the error along to the next handler in the chain. The ES handler serializes error information using its own mapping of the Elastic Common Schema. This data can accept additional user provided metadata that can assist users in searching and reporting on failures (see labels and tags settings below).

Available configurations for this handler:

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

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.

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 

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

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 

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.

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."); 
    }
}

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]:

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.

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);
        }
    }
}

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].

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   

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).

Before sending data to Elasticsearch, elasticsearch-hadoop must serialize each document into a JSON bulk entry. It is during this process that the bulk operation is determined, document metadata is extracted, and integration specific data structures are converted into JSON documents. During this process, inconsistencies with record structure can cause exceptions to be thrown during the serialization process. These errors often lead to failed tasks and halted processing.

Elasticsearch for Apache Hadoop provides an API to handle serialization errors at the record level. Error handlers for serialization are given:

There are a few default error handlers provided by the connector:

Default Handler Name: log

When this handler is invoked it logs a message containing the data structure’s toString() contents 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:

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.

Default Handler Name: es

When this handler is invoked it converts the error given into a JSON document and inserts it into an Elasticsearch index. When this indexing operation is complete, the handler can be configured to signal that the error is handled (default), or pass the error along to the next handler in the chain. The ES handler serializes error information using its own mapping of the Elastic Common Schema. This data can accept additional user provided metadata that can assist users in searching and reporting on failures (see labels and tags settings below).

Available configurations for this handler:

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

For serialization failures, the Abort on Failure built-in handler is always placed as the last error handler to catch any unhandled errors. This error handler forms the default serialization error handling behavior for elasticsearch-hadoop, which matches the behavior from previous versions.

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.data.error.handlers = log 

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

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.data.error.handlers = log 
es.write.data.error.handler.log.logger.name = SerializationErrors 

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.

Let’s say that you are streaming some unstructured data to Elasticsearch. In this scenario, your data is not fully sanitized and may contain field values that cannot be translated to JSON by the connector. You may not want to have your streaming job fail on this data, as you are potentially expecting it to contain errors. In this situation, you may want to log the data in a more comprehensive manner than to rely on the logging solution’s toString() method for your data.

Let’s write an error handler for this situation:

package org.myproject.myhandlers;

import org.elasticsearch.hadoop.handler.HandlerResult;
import org.elasticsearch.hadoop.handler.ErrorCollector;
import org.elasticsearch.hadoop.serialization.handler.write.SerializationErrorHandler;
import org.elasticsearch.hadoop.serialization.handler.write.SerializationFailure;

public class CustomLogOnError extends SerializationErrorHandler {      

    private Log logger = ???; 

    @Override
    public HandlerResult onError(SerializationFailure entry, ErrorCollector<Object> collector) throws Exception {  
        MyRecord record = (MyRecord) entry.getRecord();                             
        logger.error("Could not serialize record. " +
                "Record data : " + record.getSpecificField() + ", " + record.getOtherField(), entry.getException()); 
        return HandlerResult.HANDLED;                                               
    }
}

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

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

es.write.data.error.handler.customLog = org.myproject.myhandlers.CustomLogOnError

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

es.write.data.error.handlers = customLog
es.write.data.error.handler.customLog = org.myproject.myhandlers.CustomLogOnError

Now, your custom logging error handler will be invoked whenever a serialization failure occurs, and will instruct the connector that it is ok with ignoring those failures to continue processing.

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.handler.ErrorCollector;
import org.elasticsearch.hadoop.serialization.handler.write.SerializationErrorHandler;
import org.elasticsearch.hadoop.serialization.handler.write.SerializationFailure;

public class OutputToFileHandler extends SerializationErrorHandler { 

    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(SerializationFailure entry, ErrorCollector<Object> collector)   
    throws Exception
    {
        writer.write("Record: " + entry.getRecord().toString());
        writer.newLine();
        writer.write("Error: " + entry.getException().getMessage());
        writer.newLine();
        for (String message : entry.previousHandlerMessages()) {
            writer.write("Previous Handler: " + message);           
            writer.newLine();
        }

        return HandlerResult.PASS; 
    }

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

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.data.error.handler.[HANDLER-NAME].[PROPERTY-NAME].

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

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

Now to bring it all together:

es.write.data.error.handlers = writeFile,customLog

es.write.data.error.handler.customLog = org.myproject.myhandlers.CustomLogOnError

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

You now have a chain of handlers that writes all relevant data about the failure to a file (our writeFile handler), then logs the errors using a custom log line and ignores the error to continue processing (our customLog handler).

When reading data, the connector executes scroll requests against the configured indices and reads their contents. For each hit in a scroll search result, the connector attempts to deserialize it into an integration specific record type. When using MapReduce, this data type is either a MapWritable or Text (for raw JSON data). For an integration like Spark SQL which uses data schemas, the resulting data type is a Row object.

Elasticsearch stores documents in lucene indices. These documents can sometimes have loose definitions, or have structures that cannot be parsed into a schema-based data type, for one reason or another. Sometimes a field may be in a format that cannot be read correctly.

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

Note: Deserialization Error Handlers only allow handling of errors that occur when parsing documents from scroll responses. It may be possible that a search result can be successfully read, but is still malformed, thus causing an exception when it is used in a completely different part of the framework. This Error Handler is called from the top of the most reasonable place to handle exceptions in the scroll reading process, but this does not encapsulate all logic for each integration.

There are a few default error handlers provided by the connector:

Default Handler Name: log

When this handler is invoked it logs a message containing the JSON search hit 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:

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.

Default Handler Name: es

When this handler is invoked it converts the error given into a JSON document and inserts it into an Elasticsearch index. When this indexing operation is complete, the handler can be configured to signal that the error is handled (default), or pass the error along to the next handler in the chain. The ES handler serializes error information using its own mapping of the Elastic Common Schema. This data can accept additional user provided metadata that can assist users in searching and reporting on failures (see labels and tags settings below).

Available configurations for this handler:

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

For deserialization failures, the Abort on Failure built-in handler is always placed as the last error handler to catch any unhandled errors. This error handler alone forms the default deserialization error handling behavior for elasticsearch-hadoop, which matches the behavior from previous versions.

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.read.data.error.handlers = log 

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

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.read.data.error.handlers = log 
es.read.data.error.handler.log.logger.name = DeserializationErrors 

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.

Let’s say that you are reading a large index of log data from Elasticsearch. In this scenario, your log data is highly unstructured, and not all of its contents are critical to your process. Due to the volume of data being read, your job takes a long time to complete. In this case, you might want to replace records that cannot be read with a dummy record to mark the failure, and not interrupt your processing. The offending data should be logged and dropped.

Let’s write an error handler for this situation:

package org.myproject.myhandlers;

import org.elasticsearch.hadoop.handler.HandlerResult;
import org.elasticsearch.hadoop.handler.ErrorCollector;
import org.elasticsearch.hadoop.serialization.handler.read.DeserializationErrorHandler;
import org.elasticsearch.hadoop.serialization.handler.read.DeserializationFailure;

public class ReturnDummyHandler extends DeserializationErrorHandler { 

    private static final Logger LOGGER = ...; 
    private static final String DUMMY_RECORD = "..."; 

    @Override
    public HandlerResult onError(DeserializationFailure entry, ErrorCollector<byte[]> collector) 
    throws Exception
    {
        BufferedReader reader = new BufferedReader(new InputStreamReader(entry.getHitContents()));
        StringBuilder hitContent = new StringBuilder();
        for (String line = reader.readLine(); line != null; line = reader.readLine()) {           
            hitContent.append(line);
        }
        LOGGER.warn("Encountered malformed record during read. Replacing with dummy record. " +   
                            "Malformed Data: " + hitContent, entry.getException());
        return collector.retry(DUMMY_RECORD.getBytes());                                         
    }
}

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

In this case, lets register a handler name for our dummy record handler:

es.read.data.error.handler.returnDummy = org.myproject.myhandlers.ReturnDummyHandler

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

es.read.data.error.handlers = returnDummy
es.read.data.error.handler.returnDummy = org.myproject.myhandlers.ReturnDummyHandler

Now, your dummy data error handler will be invoked whenever a deserialization failure occurs, and will instruct the connector to use your provided dummy record instead of the malformed data.

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.handler.ErrorCollector;
import org.elasticsearch.hadoop.serialization.handler.read.DeserializationErrorHandler;
import org.elasticsearch.hadoop.serialization.handler.read.DeserializationFailure;

public class ReturnDummyAndLogToFileHandler extends DeserializationErrorHandler { 

    private static final String DUMMY_RECORD = "...";

    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(DeserializationFailure entry, ErrorCollector<byte[]> collector)   
    throws Exception
    {
        BufferedReader reader = new BufferedReader(new InputStreamReader(entry.getHitContents()));
        StringBuilder hitContent = new StringBuilder();
        for (String line = reader.readLine(); line != null; line = reader.readLine()) {           
            hitContent.append(line);
        }

        writer.write("Error: " + entry.getException().getMessage());
        writer.newLine();
        for (String message : entry.previousHandlerMessages()) {
            writer.write("Previous Handler: " + message);           
            writer.newLine();
        }
        writer.write("Entry: ");
        writer.newLine();
        writer.write(hitContent.toString());
        writer.newLine();

        return collector.retry(DUMMY_RECORD.getBytes());            
    }

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

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

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

es.read.data.error.handler.writeFile = org.myproject.myhandlers.ReturnDummyAndLogToFileHandler   
es.read.data.error.handler.writeFile.filename = /path/to/some/output/file   

Now to bring it all together with the previous example:

es.read.data.error.handlers = writeFile
es.read.data.error.handler.writeFile = org.myproject.myhandlers.ReturnDummyAndLogToFileHandler
es.read.data.error.handler.writeFile.filename = /path/to/some/output/file

You now have a handler that retries replaces malformed data with dummy records, then logs those malformed records along with their error information by writing them out to a custom file.