Thanks Bryan !!

Connection is in ESTBLISHED state on on end and completely missing at other end (in another dc).

Yes, we can revisit TCP tuning.But the problem is node specific. So not sure whether tuning is the culprit.

Thanks
Anuj

Sent from Yahoo Mail on Android


From:"Bryan Cheng" <bryan@blockcypher.com>
Date:Wed, 18 Nov, 2015 at 2:04 am
Subject:Re: Repair Hangs while requesting Merkle Trees

Ah OK, might have misunderstood you. Streaming socket should not be in play during merkle tree generation (validation compaction). They may come in play during merkle tree exchange- that I'm not sure about. You can read a bit more here: https://issues.apache.org/jira/browse/CASSANDRA-8611.

Regardless, you should have it set- 1 hr is usually a good conservative estimate, but you can go much lower safely.

What state is the connection on that only shows on one side? Is it ESTABLISHED, or something like CLOSE_WAIT?

Here's a good place to start for tuning, though it doesn't have as much about network tuning: https://tobert.github.io/pages/als-cassandra-21-tuning-guide.html. More generally, TCP tuning usually revolves around a balance between latency and bandwidth. Over long connections (we're talking 10s of ms, instead of the sub 1ms you usually see in a good dc network), your expectations will shift greatly. Stuff like NODELAY on tcp is very nice for cutting your latencies when you're inside a DC, but will generate lots of small packets that will hurt your bandwidth over longer connections due to the need to wait for acks. otc_coalescing_strategy is on a similar vein, bundling together nearby messages to trade latency for throughput. You'll also probably want to tune your tcp buffers and window sizes, since that determines how much data can be in-flight between acknowledgements, and the default size is pitiful for any decent  network size. Google around for TCP tuning/buffer tuning and you should find some good resources.

On Mon, Nov 16, 2015 at 5:23 PM, Anuj Wadehra <anujw_2003@yahoo.co.in> wrote:
Hi Bryan,

Thanks for the reply !!
I didnt mean streaming_socket_tomeout_in_ms. I meant when you run netstats (Linux cmnd) on  node A in DC1, you will notice that there is connection in established state with node B in DC2. But when you run netstats on node B, you wont find any connection with node A. Such connections are there across dc? Is it a problem.

We havent set streaming_socket_timeout_in_ms which I know must be set. But I am not  sure wtheher setting this property has any impact on merkle tree requests. I thought its valid for data streaming if some mismatch is found and data needs to be streamed.Please confirm. Whats the value you use for streaming socket timeout?

Morever, if socket timeout is the issue, that should happen on other nodes too...repair is not running on just one node, as merkle tree request is getting lost n not transmitted to one or more nodes in remote dc.

I am not sure about exact distance. But they are connected with a very high speed 10gbps link.

When you say different TCP stack tuning..do u have any document/blog/link describing recommendations for multi Dc Cassandra setup?  Can you elaborate what all settings need to be different? 


Thanks
Anuj








Sent from Yahoo Mail on Android


From:"Bryan Cheng" <bryan@blockcypher.com>
Date:Tue, 17 Nov, 2015 at 5:54 am

Subject:Re: Repair Hangs while requesting Merkle Trees

Hi Anuj,

Did you mean streaming_socket_timeout_in_ms? If not, then you definitely want that set. Even the best network connections will break occasionally, and in Cassandra < 2.1.10 (I believe) this would leave those connections hanging indefinitely on one end.

How far away are your two DC's from a network perspective, out of curiosity? You'll almost certainly be doing different TCP stack tuning for cross-DC, notably your buffer sizes, window params, cassandra-specific stuff like otc_coalescing_strategy, inter_dc_tcp_nodelay, etc.

On Sat, Nov 14, 2015 at 10:35 AM, Anuj Wadehra <anujw_2003@yahoo.co.in> wrote:
One more observation.We observed that there are few TCP connections which node shows as Established but when we go to node at other end,connection is not there. They are called "phantom" connections I guess. Can this be a possible cause?

Thanks
Anuj

Sent from Yahoo Mail on Android


From:"Anuj Wadehra" <anujw_2003@yahoo.co.in>
Date:Sat, 14 Nov, 2015 at 11:59 pm

Subject:Re: Repair Hangs while requesting Merkle Trees

Thanks Daemeon !!

I wil capture the output of netstats and share in next few days. We were thinking of taking tcp dumps also. If its a network issue and increasing request timeout worked, not sure how Cassandra is dropping messages based on timeout.Repair messages are non droppable and not supposed to be timedout.

2 of the 3 nodes in the DC are able to complete repair without any issue. Just one node is problematic.

I also observed frequent messages in logs of other nodes which say that hints replay timedout..and the node where hints were being replayed is always a remote dc node. Is it related some how?

Thanks
Anuj

Sent from Yahoo Mail on Android


From:"daemeon reiydelle" <daemeonr@gmail.com>
Date:Thu, 12 Nov, 2015 at 10:34 am
Subject:Re: Repair Hangs while requesting Merkle Trees


Have you checked the network statistics on that machine? (netstats -tas) while attempting to repair ... if netstats show ANY issues you have a problem. If you can put the command in a loop running every 60 seconds for maybe 15 minutes and post back?

Out of curiousity, how many remote DC nodes are getting successfully repaired?


.......
“Life should not be a journey to the grave with the intention of arriving safely in a
pretty and well preserved body, but rather to skid in broadside in a cloud of smoke,
thoroughly used up, totally worn out, and loudly proclaiming “Wow! What a Ride!”

- Hunter Thompson

Daemeon C.M. Reiydelle
USA (+1) 415.501.0198
London (+44) (0) 20 8144 9872


On Wed, Nov 11, 2015 at 1:06 PM, Anuj Wadehra <anujw_2003@yahoo.co.in> wrote:
Hi,

we are using 2.0.14. We have 2 DCs at remote locations with 10GBps connectivity.We are able to complete repair (-par -pr) on 5 nodes. On only one node in DC2, we are unable to complete repair as it always hangs. Node sends Merkle Tree requests, but one or more nodes in DC1 (remote) never show that they sent the merkle tree reply to requesting node.
Repair hangs infinitely.

After increasing request_timeout_in_ms on affected node, we were able to successfully run repair on one of the two occassions.

Any comments, why this is happening on just one node? In OutboundTcpConnection.java,  when isTimeOut method always returns false for non-droppable verb such as Merkle Tree Request(verb=REPAIR_MESSAGE),why increasing request timeout solved problem on one occasion ?


Thanks
Anuj Wadehra



On Thursday, 12 November 2015 2:35 AM, Anuj Wadehra <anujw_2003@yahoo.co.in> wrote:


Hi,

We have 2 DCs at remote locations with 10GBps connectivity.We are able to complete repair (-par -pr) on 5 nodes. On only one node in DC2, we are unable to complete repair as it always hangs. Node sends Merkle Tree requests, but one or more nodes in DC1 (remote) never show that they sent the merkle tree reply to requesting node.
Repair hangs infinitely.

After increasing request_timeout_in_ms on affected node, we were able to successfully run repair on one of the two occassions.

Any comments, why this is happening on just one node? In OutboundTcpConnection.java,  when isTimeOut method always returns false for non-droppable verb such as Merkle Tree Request(verb=REPAIR_MESSAGE),why increasing request timeout solved problem on one occasion ?


Thanks
Anuj Wadehra