segmentation fault

[aerospike]

segmentation fault

[GayathriKaliyamoorthy]

Hi Hamper,
I am not able to reproduce the issue in house.
Can you give input on the kind of load that generates this segfault, and the client machine configuration on which this high load is being run?

Thanks.

[Hamper]

this test script can reproduce error: https://dl.dropboxusercontent.com/u/12199274/github/test.tar.gz

client machine: Intel® Core™ i7-4770 Quadcore Haswell, 32 GB DDR3 RAM, Ubuntu 12.04

[GayathriKaliyamoorthy]

Hamper,
Thanks for providing us with the scripts to reproduce the problem.
I ran the script in ubuntu, debian7 and centos6 for more than 4 hours. But could not reproduce the same segfault. The stacktrace you reported shows some memory allocation failure. We have fixed a bunch of memory leaks and other problems, released the latest version to npm. Could you give a shot now and see if you land in the same problem again?

Thanks.

[Hamper]

I still can reproduce with this script (ubuntu 14.10, nodejs 0.10.35, aerospikeclient 1.0.26, aerospike server 3.4.1 with default settings), also node main.js -P 1 -I 10 -O 100000 from benchmark reproduces error

[courtneycouch]

We're also getting sporadic segmentation faults under high loads.

[courtneycouch]

We set up totally fresh servers on ec2 (r3.xlarge instances) running Ubuntu and only the bare minimum installed to get the nodejs client, and the bare minimum to get aerospike installed to verify there wasn't anything odd conflicting.

Both are: Ubuntu 14.04.1 LTS and Node.js version: v0.10.36

In case it impacts anything the aerospike server was initialized with 10m keys. Doesn't seem like the aerospike server would have any impact here as it's a client error though.

We're considering moving some of our higher load data over but this would be accessed through the nodejs client.. as a result it really needs to be rock solid under heavy highly concurrent load.

Using this we can get the faults pretty consistently. Sometimes it'll nearly finish, other times the fault happens right away. The Aerospike server is remote to the client as well.

Obviously this is hugely stripped down.. just wanted the bare minimum to reproduce it there.

var async = require('async');
var aerospike = require('aerospike');
var status = aerospike.status;

var client = aerospike.client({
    hosts: [ { addr: '10.10.10.220', port: 3000 } ]
});

var key = {
  ns: 'store_disk_ebs',
  set: 'benchmark',
  key: 'keytest'
};

var iteration_count = 100000;
var concurrency_count = 10;

//connect to server
client.connect(function(err, client) {

  //verify connected ok
  if (err.code != status.AEROSPIKE_OK) throw 'failed connection';

  //loop iteration_count number of times in a series
  async.timesSeries(iteration_count, function(n, iteration) {

    //print iteration count every 1000 iterations
    if (n % 1000 === 0) console.log('iteration: ' + n);

    //for each iteration run this command concurrency_count times in parallel
    async.times(concurrency_count, function(n, concurrent) {

      //get the key
      client.get(key, function() {
        //we don't need to do anything with the result. just complete this command
        concurrent();
      });

    }, iteration); //after the concurrency_count commands are complete, start a new iteration

  });
});

[GayathriKaliyamoorthy]

Thanks for your patience.

We have identified the issue. There was a problem with the way we were handling buffers in V8 layer. We fixed that and made an official release to npm. The latest version is 1.0.28. Please use the latest and give us your feedback.

Thanks.

[vivekkrbajpai]

Hey Gayathri,

It seems the version 1.0.28 also contains this error. I am getting these things in amazon linux log messages.

[ 1514.330637] node[21522]: segfault at 4 ip 00007fea72009833 sp 00007fea6bffea88 error 6 in libc-2.17.so[7fea71eda000+19b000]
Feb 9 16:47:15 ip-xx-xxx-xx-xx kernel: [ 1514.375729] node[21531]: segfault at 4 ip 00007fe77003d833 sp 00007fe76e059a88 error 6 in libc-2.17.so[7fe76ff0e000+19b000]
Feb 9 16:47:15 ip-xx-xxx-xx-xx kernel: [ 1514.381353] node[21538]: segfault at 4 ip 00007f9bdc501833 sp 00007f9bd6be2a88 error 6 in libc-2.17.so[7f9bdc3d2000+19b000]
Feb 9 16:47:15 ip-xx-xxx-xx-xx kernel: [ 1514.388211]
Feb 9 16:47:16 ip-xx-xxx-xx-xx kernel: [ 1515.116007] node[21566]: segfault at 4 ip 00007f16308f9833 sp 00007f162affca88 error 6 in libc-2.17.so[7f16307ca000+19b000]
Feb 9 16:47:17 ip-xx-xxx-xx-xx kernel: [ 1516.078188] node[21595]: segfault at 4 ip 00007eff66f56833 sp 00007eff64f72a88 error 6 in libc-2.17.so[7eff66e27000+19b000]
Feb 9 16:47:17 ip-xx-xxx-xx-xx kernel: [ 1516.672639] node[21599]: segfault at 4 ip 00007f13bb000833 sp 00007f13b901ca88 error 6 in libc-2.17.so[7f13baed1000+19b000]
Feb 9 16:47:17 ip-xx-xxx-xx-xx kernel: [ 1516.694516] node[21591]: segfault at 4 ip 00007fc4cfe8e833 sp 00007fc4ccea8a88 error 6 in libc-2.17.so[7fc4cfd5f000+19b000]
Feb 9 16:47:17 ip-xx-xxx-xx-xx kernel: [ 1516.901260] node[21631]: segfault at 4 ip 00007fa381e3c833 sp 00007fa37b7fda88 error 6 in libc-2.17.so[7fa381d0d000+19b000]
Feb 9 16:47:18 ip-xx-xxx-xx-xx kernel: [ 1517.719136] node[21643]: segfault at 4 ip 00007ff011a87833 sp 00007ff00a7fba88 error 6 in libc-2.17.so[7ff011958000+19b000]

Could you please check.

[ryanwitt]

@vivekkrbajpai are you using batchGet()? I'm getting segfaults in this function on 1.0.28 and not with get()

[GayathriKaliyamoorthy]

@ryanwitt Thanks for the input regarding batch_get, I'll try reproducing the segfault with batch_get and root cause the issue.

Thanks

[vivekkrbajpai]

@GayathriKaliyamoorthy @ryanwitt yeah i am using batchGet(). During high concurrent load it produces segfault.

[GayathriKaliyamoorthy]

@vivekkrbajpai @ryanwitt I have identified a loop hole in batchGet logic, where if one of the keys in the batchKeys is corrupted or not constructed properly the driver segfaults. Can you confirm that, your application always sends a well constructed batchKeys to aerospike nodejs driver.

And also I tried reproducing a segfault under heavy load, but could not reproduce. Could you give a sample code snippet so that I can work on reproducing the segfault under high load.

Thanks

[vivekkrbajpai]

Yups my keys are well formed and not corrupted or empty.
On Feb 10, 2015 7:02 PM, "Gayathri" [email protected] wrote:

@vivekkrbajpai https://github.com/vivekkrbajpai @ryanwitt
https://github.com/ryanwitt I have identified a loop hole in batchGet
logic, where if one of the keys in the batchKeys is corrupted or not
constructed properly the driver segfaults. Can you confirm that, your
application always sends a well constructed batchKeys to aerospike nodejs
driver.

And also I tried reproducing a segfault under heavy load, but could not
reproduce. Could you give a sample code snippet so that I can work on
reproducing the segfault under high load.

Thanks

—
Reply to this email directly or view it on GitHub
#26 (comment)
.

[ryanwitt]

They're all of the form:

var keys = [
 { ns: 'namespace', set: 'some_set', key: 'some, possibly very long key' },
 ...
];

Is there any limit to the length of the key itself?

[GayathriKaliyamoorthy]

Could you give an approximate size of each batch requests?

Thanks

[vivekkrbajpai]

Its about 5 to 10 in each batch.

On Wed, Feb 11, 2015 at 2:57 PM, Gayathri [email protected] wrote:

Could you give an approximate size of each batch requests?

Thanks

—
Reply to this email directly or view it on GitHub
#26 (comment)
.

[GayathriKaliyamoorthy]

@Hamper @courtneycouch One of our customers had reported back saying after the fix, the driver ran without any segfault for 19 continuous hours, could you also confirm this please?

Thanks

[GayathriKaliyamoorthy]

@vivekkrbajpai @ryanwitt could you please open another issue stating segfault in batchGet API, it will be easy for us to track the issues.

Thanks

[vivekkrbajpai]

Sure

[Hamper]

It's work, Thanks.