You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@flume.apache.org by Paul Merry <pa...@bbc.co.uk> on 2014/02/05 09:34:08 UTC
Channel management: messages that will never be delivered
Hi,
We are using an Elasticsearch sink and have seen a file channel filling with messages that will never be delivered as the format of the message is incompatible with Elasticsearch itself.
Example message from Flume logs:
24 Jan 2014 08:14:55,173 ERROR [SinkRunner-PollingRunner-DefaultSinkProcessor]
(org.apache.flume.SinkRunner$PollingRunner.run:160) - Unable to deliver event.
Exception follows.
org.elasticsearch.indices.InvalidIndexNameException: [UpperCase-2014-01-23]
Invalid index name [UpperCase-2014-01-23], must be lowercase
In this case the index name comes from a header so we have a workaround using a multiplexing channel selector to detect and re-route messages based on headers of this format.
To clean up the channel this time we removed the data and checkpoint directories, which is not ideal as we probably lost other messages in doing this.
We are wary of similar situations occurring in future for messages that we can't detect and divert in advance and so have a few questions:
- What would be the recommended handling of this situation?
- Is it possible to clear just these messages from the channel or does the whole channel have to be deleted ?
- Is there a way that we can divert these messages to another channel (dead letter / invalid message style) ? Noting that they are not known to be problematic until after an attempt is made to deliver them from the sink
- What happens to other messages in a batch with a bad message ? Will they also be stuck forever or will they be taken in another batch ?
Thanks,
Paul.
----------------------------
http://www.bbc.co.uk
This e-mail (and any attachments) is confidential and may contain personal views which are not the views of the BBC unless specifically stated.
If you have received it in error, please delete it from your system.
Do not use, copy or disclose the information in any way nor act in reliance on it and notify the sender immediately.
Please note that the BBC monitors e-mails sent or received.
Further communication will signify your consent to this.
---------------------
RE: Channel management: messages that will never be delivered
Posted by Paul Merry <pa...@bbc.co.uk>.
Thanks Matt, that seems like an approach we could also take, good to have a clearer view of where the limitations of the current system are.
________________________________
From: Kenison, Matt [Matt.Kenison@disney.com]
Sent: 06 February 2014 21:11
To: user@flume.apache.org
Subject: Re: Channel management: messages that will never be delivered
It's possible, but not easy to do. In our application, the individual custom sinks know which exceptions can rollback or are unrecoverable, but this doesn't work for the built-in sinks, and it doesn't take into account unexpected failures. So, we control it manually with a JMX flag and by subclassing BasicTransactionSemantics. When the flag is set, and the transaction tries to rollback, it performs a commit instead (and directs the messages to a failure channel). When a transaction is successful, the flag is reset.
It's not the prettiest solution, but isn't a hack. It requires subclassing the channel to provide a custom transaction, and overriding the default transaction behavior. Flume really doesn't make it easy to extend behavior of any of the standard components.
From: Paul Merry <pa...@bbc.co.uk>>
Reply-To: "user@flume.apache.org<ma...@flume.apache.org>" <us...@flume.apache.org>>
Date: Thursday, February 6, 2014 12:28 AM
To: "user@flume.apache.org<ma...@flume.apache.org>" <us...@flume.apache.org>>
Subject: RE: Channel management: messages that will never be delivered
Thanks for the suggestion Ed, it's definitely something we could look at.
I did find this ticket https://issues.apache.org/jira/browse/FLUME-2140 and the linked discussion thread http://flume.markmail.org/thread/y3cks6hdgof3kxu6#query:+page:1+mid:rx3zm53t4dhmqskk+state:results
There are some suggestions for work arounds there, probably the use of failiover sink is most relevant but I'd be concerned for what might happen to legitimate messages in a situation where there is downtime or connection issues with the endpoint. It seems we'd loose the correct channel retry logic (in that scenario) and end up with messages that would need replaying.
If there isn't much to add on the handling of 'bad messages' can anyone inform on the handling of other messages in a batch with one or more of these messages that will never deliver. Will they also not make it to their destination or will they get rebatched?
Also keen for anyone with an idea for how to clear these messages from a channel once they are stuck, as the directory deletion can take good messages down too.
- Paul
________________________________
From: ed [edorsey@gmail.com<ma...@gmail.com>]
Sent: 05 February 2014 23:12
To: user@flume.apache.org<ma...@flume.apache.org>
Subject: Re: Channel management: messages that will never be delivered
Hi Paul,
Not sure if this would work for you but if you can error check prior to the events reaching Elasticsearch you can handle this by writing a custom Interceptor that validates your events. You can do more robust error checking here than you can by just relying on the already existing event header fields as you'll have full access to the event header and body. Within the interceptor, if the event is not compatible with Elasticsearch you can add a boolean flag to the header of the event like "hasError". Then you can route any events that have an error to a different "error" channel using a multiplexing selector by checking for the hasError flag. The error channel can either be connected to a NullSink or a FileRoll sink if you want to preserve the improperly formatted events.
We've only used the memory channel so far so I'm afraid I can't comment on the file channel specific questions you have. Hopefully someone on the list with some more experience there can chime in.
Best,
Ed
On Wed, Feb 5, 2014 at 5:34 PM, Paul Merry <pa...@bbc.co.uk>> wrote:
Hi,
We are using an Elasticsearch sink and have seen a file channel filling with messages that will never be delivered as the format of the message is incompatible with Elasticsearch itself.
Example message from Flume logs:
24 Jan 2014 08:14:55,173 ERROR [SinkRunner-PollingRunner-DefaultSinkProcessor]
(org.apache.flume.SinkRunner$PollingRunner.run:160) - Unable to deliver event.
Exception follows.
org.elasticsearch.indices.InvalidIndexNameException: [UpperCase-2014-01-23]
Invalid index name [UpperCase-2014-01-23], must be lowercase
In this case the index name comes from a header so we have a workaround using a multiplexing channel selector to detect and re-route messages based on headers of this format.
To clean up the channel this time we removed the data and checkpoint directories, which is not ideal as we probably lost other messages in doing this.
We are wary of similar situations occurring in future for messages that we can't detect and divert in advance and so have a few questions:
- What would be the recommended handling of this situation?
- Is it possible to clear just these messages from the channel or does the whole channel have to be deleted ?
- Is there a way that we can divert these messages to another channel (dead letter / invalid message style) ? Noting that they are not known to be problematic until after an attempt is made to deliver them from the sink
- What happens to other messages in a batch with a bad message ? Will they also be stuck forever or will they be taken in another batch ?
Thanks,
Paul.
----------------------------
http://www.bbc.co.uk
This e-mail (and any attachments) is confidential and may contain personal views which are not the views of the BBC unless specifically stated.
If you have received it in error, please delete it from your system.
Do not use, copy or disclose the information in any way nor act in reliance on it and notify the sender immediately.
Please note that the BBC monitors e-mails sent or received.
Further communication will signify your consent to this.
---------------------
----------------------------
http://www.bbc.co.uk
This e-mail (and any attachments) is confidential and may contain personal views which are not the views of the BBC unless specifically stated.
If you have received it in error, please delete it from your system.
Do not use, copy or disclose the information in any way nor act in reliance on it and notify the sender immediately.
Please note that the BBC monitors e-mails sent or received.
Further communication will signify your consent to this.
---------------------
----------------------------
http://www.bbc.co.uk
This e-mail (and any attachments) is confidential and may contain personal views which are not the views of the BBC unless specifically stated.
If you have received it in error, please delete it from your system.
Do not use, copy or disclose the information in any way nor act in reliance on it and notify the sender immediately.
Please note that the BBC monitors e-mails sent or received.
Further communication will signify your consent to this.
---------------------
Re: Channel management: messages that will never be delivered
Posted by "Kenison, Matt" <Ma...@disney.com>.
It's possible, but not easy to do. In our application, the individual custom sinks know which exceptions can rollback or are unrecoverable, but this doesn't work for the built-in sinks, and it doesn't take into account unexpected failures. So, we control it manually with a JMX flag and by subclassing BasicTransactionSemantics. When the flag is set, and the transaction tries to rollback, it performs a commit instead (and directs the messages to a failure channel). When a transaction is successful, the flag is reset.
It's not the prettiest solution, but isn't a hack. It requires subclassing the channel to provide a custom transaction, and overriding the default transaction behavior. Flume really doesn't make it easy to extend behavior of any of the standard components.
From: Paul Merry <pa...@bbc.co.uk>>
Reply-To: "user@flume.apache.org<ma...@flume.apache.org>" <us...@flume.apache.org>>
Date: Thursday, February 6, 2014 12:28 AM
To: "user@flume.apache.org<ma...@flume.apache.org>" <us...@flume.apache.org>>
Subject: RE: Channel management: messages that will never be delivered
Thanks for the suggestion Ed, it's definitely something we could look at.
I did find this ticket https://issues.apache.org/jira/browse/FLUME-2140 and the linked discussion thread http://flume.markmail.org/thread/y3cks6hdgof3kxu6#query:+page:1+mid:rx3zm53t4dhmqskk+state:results
There are some suggestions for work arounds there, probably the use of failiover sink is most relevant but I'd be concerned for what might happen to legitimate messages in a situation where there is downtime or connection issues with the endpoint. It seems we'd loose the correct channel retry logic (in that scenario) and end up with messages that would need replaying.
If there isn't much to add on the handling of 'bad messages' can anyone inform on the handling of other messages in a batch with one or more of these messages that will never deliver. Will they also not make it to their destination or will they get rebatched?
Also keen for anyone with an idea for how to clear these messages from a channel once they are stuck, as the directory deletion can take good messages down too.
- Paul
________________________________
From: ed [edorsey@gmail.com<ma...@gmail.com>]
Sent: 05 February 2014 23:12
To: user@flume.apache.org<ma...@flume.apache.org>
Subject: Re: Channel management: messages that will never be delivered
Hi Paul,
Not sure if this would work for you but if you can error check prior to the events reaching Elasticsearch you can handle this by writing a custom Interceptor that validates your events. You can do more robust error checking here than you can by just relying on the already existing event header fields as you'll have full access to the event header and body. Within the interceptor, if the event is not compatible with Elasticsearch you can add a boolean flag to the header of the event like "hasError". Then you can route any events that have an error to a different "error" channel using a multiplexing selector by checking for the hasError flag. The error channel can either be connected to a NullSink or a FileRoll sink if you want to preserve the improperly formatted events.
We've only used the memory channel so far so I'm afraid I can't comment on the file channel specific questions you have. Hopefully someone on the list with some more experience there can chime in.
Best,
Ed
On Wed, Feb 5, 2014 at 5:34 PM, Paul Merry <pa...@bbc.co.uk>> wrote:
Hi,
We are using an Elasticsearch sink and have seen a file channel filling with messages that will never be delivered as the format of the message is incompatible with Elasticsearch itself.
Example message from Flume logs:
24 Jan 2014 08:14:55,173 ERROR [SinkRunner-PollingRunner-DefaultSinkProcessor]
(org.apache.flume.SinkRunner$PollingRunner.run:160) - Unable to deliver event.
Exception follows.
org.elasticsearch.indices.InvalidIndexNameException: [UpperCase-2014-01-23]
Invalid index name [UpperCase-2014-01-23], must be lowercase
In this case the index name comes from a header so we have a workaround using a multiplexing channel selector to detect and re-route messages based on headers of this format.
To clean up the channel this time we removed the data and checkpoint directories, which is not ideal as we probably lost other messages in doing this.
We are wary of similar situations occurring in future for messages that we can't detect and divert in advance and so have a few questions:
- What would be the recommended handling of this situation?
- Is it possible to clear just these messages from the channel or does the whole channel have to be deleted ?
- Is there a way that we can divert these messages to another channel (dead letter / invalid message style) ? Noting that they are not known to be problematic until after an attempt is made to deliver them from the sink
- What happens to other messages in a batch with a bad message ? Will they also be stuck forever or will they be taken in another batch ?
Thanks,
Paul.
----------------------------
http://www.bbc.co.uk
This e-mail (and any attachments) is confidential and may contain personal views which are not the views of the BBC unless specifically stated.
If you have received it in error, please delete it from your system.
Do not use, copy or disclose the information in any way nor act in reliance on it and notify the sender immediately.
Please note that the BBC monitors e-mails sent or received.
Further communication will signify your consent to this.
---------------------
----------------------------
http://www.bbc.co.uk
This e-mail (and any attachments) is confidential and may contain personal views which are not the views of the BBC unless specifically stated.
If you have received it in error, please delete it from your system.
Do not use, copy or disclose the information in any way nor act in reliance on it and notify the sender immediately.
Please note that the BBC monitors e-mails sent or received.
Further communication will signify your consent to this.
---------------------
RE: Channel management: messages that will never be delivered
Posted by Paul Merry <pa...@bbc.co.uk>.
Thanks for the suggestion Ed, it's definitely something we could look at.
I did find this ticket https://issues.apache.org/jira/browse/FLUME-2140 and the linked discussion thread http://flume.markmail.org/thread/y3cks6hdgof3kxu6#query:+page:1+mid:rx3zm53t4dhmqskk+state:results
There are some suggestions for work arounds there, probably the use of failiover sink is most relevant but I'd be concerned for what might happen to legitimate messages in a situation where there is downtime or connection issues with the endpoint. It seems we'd loose the correct channel retry logic (in that scenario) and end up with messages that would need replaying.
If there isn't much to add on the handling of 'bad messages' can anyone inform on the handling of other messages in a batch with one or more of these messages that will never deliver. Will they also not make it to their destination or will they get rebatched?
Also keen for anyone with an idea for how to clear these messages from a channel once they are stuck, as the directory deletion can take good messages down too.
- Paul
________________________________
From: ed [edorsey@gmail.com]
Sent: 05 February 2014 23:12
To: user@flume.apache.org
Subject: Re: Channel management: messages that will never be delivered
Hi Paul,
Not sure if this would work for you but if you can error check prior to the events reaching Elasticsearch you can handle this by writing a custom Interceptor that validates your events. You can do more robust error checking here than you can by just relying on the already existing event header fields as you'll have full access to the event header and body. Within the interceptor, if the event is not compatible with Elasticsearch you can add a boolean flag to the header of the event like "hasError". Then you can route any events that have an error to a different "error" channel using a multiplexing selector by checking for the hasError flag. The error channel can either be connected to a NullSink or a FileRoll sink if you want to preserve the improperly formatted events.
We've only used the memory channel so far so I'm afraid I can't comment on the file channel specific questions you have. Hopefully someone on the list with some more experience there can chime in.
Best,
Ed
On Wed, Feb 5, 2014 at 5:34 PM, Paul Merry <pa...@bbc.co.uk>> wrote:
Hi,
We are using an Elasticsearch sink and have seen a file channel filling with messages that will never be delivered as the format of the message is incompatible with Elasticsearch itself.
Example message from Flume logs:
24 Jan 2014 08:14:55,173 ERROR [SinkRunner-PollingRunner-DefaultSinkProcessor]
(org.apache.flume.SinkRunner$PollingRunner.run:160) - Unable to deliver event.
Exception follows.
org.elasticsearch.indices.InvalidIndexNameException: [UpperCase-2014-01-23]
Invalid index name [UpperCase-2014-01-23], must be lowercase
In this case the index name comes from a header so we have a workaround using a multiplexing channel selector to detect and re-route messages based on headers of this format.
To clean up the channel this time we removed the data and checkpoint directories, which is not ideal as we probably lost other messages in doing this.
We are wary of similar situations occurring in future for messages that we can't detect and divert in advance and so have a few questions:
- What would be the recommended handling of this situation?
- Is it possible to clear just these messages from the channel or does the whole channel have to be deleted ?
- Is there a way that we can divert these messages to another channel (dead letter / invalid message style) ? Noting that they are not known to be problematic until after an attempt is made to deliver them from the sink
- What happens to other messages in a batch with a bad message ? Will they also be stuck forever or will they be taken in another batch ?
Thanks,
Paul.
----------------------------
http://www.bbc.co.uk
This e-mail (and any attachments) is confidential and may contain personal views which are not the views of the BBC unless specifically stated.
If you have received it in error, please delete it from your system.
Do not use, copy or disclose the information in any way nor act in reliance on it and notify the sender immediately.
Please note that the BBC monitors e-mails sent or received.
Further communication will signify your consent to this.
---------------------
----------------------------
http://www.bbc.co.uk
This e-mail (and any attachments) is confidential and may contain personal views which are not the views of the BBC unless specifically stated.
If you have received it in error, please delete it from your system.
Do not use, copy or disclose the information in any way nor act in reliance on it and notify the sender immediately.
Please note that the BBC monitors e-mails sent or received.
Further communication will signify your consent to this.
---------------------
Re: Channel management: messages that will never be delivered
Posted by ed <ed...@gmail.com>.
Hi Paul,
Not sure if this would work for you but if you can error check prior to the
events reaching Elasticsearch you can handle this by writing a custom
Interceptor that validates your events. You can do more robust error
checking here than you can by just relying on the already existing event
header fields as you'll have full access to the event header and body.
Within the interceptor, if the event is not compatible with Elasticsearch
you can add a boolean flag to the header of the event like "hasError".
Then you can route any events that have an error to a different "error"
channel using a multiplexing selector by checking for the hasError flag.
The error channel can either be connected to a NullSink or a FileRoll sink
if you want to preserve the improperly formatted events.
We've only used the memory channel so far so I'm afraid I can't comment on
the file channel specific questions you have. Hopefully someone on the
list with some more experience there can chime in.
Best,
Ed
On Wed, Feb 5, 2014 at 5:34 PM, Paul Merry <pa...@bbc.co.uk> wrote:
> Hi,
>
> We are using an Elasticsearch sink and have seen a file channel filling
> with messages that will never be delivered as the format of the message is
> incompatible with Elasticsearch itself.
>
> Example message from Flume logs:
>
> 24 Jan 2014 08:14:55,173 ERROR [SinkRunner-PollingRunner-DefaultSinkProcessor]
> (org.apache.flume.SinkRunner$PollingRunner.run:160) - Unable to deliver event.
> Exception follows.
> org.elasticsearch.indices.InvalidIndexNameException: [UpperCase-2014-01-23]
> Invalid index name [UpperCase-2014-01-23], must be lowercase
>
>
> In this case the index name comes from a header so we have a workaround
> using a multiplexing channel selector to detect and re-route messages based
> on headers of this format.
>
> To clean up the channel this time we removed the data and checkpoint
> directories, which is not ideal as we probably lost other messages in doing
> this.
>
> We are wary of similar situations occurring in future for messages that we
> can't detect and divert in advance and so have a few questions:
>
> - What would be the recommended handling of this situation?
>
> - Is it possible to clear just these messages from the channel or does the
> whole channel have to be deleted ?
>
> - Is there a way that we can divert these messages to another channel
> (dead letter / invalid message style) ? Noting that they are not known to
> be problematic until after an attempt is made to deliver them from the sink
>
> - What happens to other messages in a batch with a bad message ? Will they
> also be stuck forever or will they be taken in another batch ?
>
>
> Thanks,
>
> Paul.
>
>
>
> ----------------------------
>
> http://www.bbc.co.uk
> This e-mail (and any attachments) is confidential and may contain personal
> views which are not the views of the BBC unless specifically stated.
> If you have received it in error, please delete it from your system.
> Do not use, copy or disclose the information in any way nor act in
> reliance on it and notify the sender immediately.
> Please note that the BBC monitors e-mails sent or received.
> Further communication will signify your consent to this.
>
> ---------------------
>