Using cloud-tpu fails with google ai-platform train api - google-cloud-ml

I've been successfully using ai-platform train api with tensor2tensor and cloud-tpu backend until several days back,
but it seems like something has changed and I can't get it to work since last week.
The differences I see in logs between working/non-working are '_master' and '_evaluation_master' from config.
Last successful log of train api shows something like below.
Using config: {
'_model_dir':...,
....,
'_master': 'grpc://10.228.38.186:8470',
'_evaluation_master': 'grpc://10.228.38.186:8470',
...
'_cluster': None, 'use_tpu': True
}
However, the logs I see since last week are as follows.
Using config: {
'_model_dir': ...,
'_master': 'cmle-training-2190487948974557758-tpu',
'_evaluation_master': 'cmle-training-2190487948974557758-tpu',
...,
'_cluster': None, 'use_tpu': True
}
Then, tensorflow tries to connect tpu by host name, which eventually fails and the process stops.
Not found: No session factory registered for the given session options:
{
target: "cmle-training-4208055151697798232-tpu"
config: operation_timeout_in_ms: 300000
}
Registered factories are {DIRECT_SESSION, GRPC_SESSION}.
Same code is used for both experiments.
If anybody has faced similar issue, please guide me through this. Thanks!

Related

Errors connecting to AWS Keyspaces using a lambda layer

Intermittently getting the following error when connecting to an AWS keyspace using a lambda layer
All host(s) tried for query failed. First host tried, 3.248.244.53:9142: Host considered as DOWN. See innerErrors.
I am trying to query a table in a keyspace using a nodejs lambda function as follows:
import cassandra from 'cassandra-driver';
import fs from 'fs';
export default class AmazonKeyspace {
tpmsClient = null;
constructor () {
let auth = new cassandra.auth.PlainTextAuthProvider('cass-user-at-xxxxxxxxxx', 'zzzzzzzzz');
let sslOptions1 = {
ca: [ fs.readFileSync('/opt/utils/AmazonRootCA1.pem', 'utf-8')],
host: 'cassandra.eu-west-1.amazonaws.com',
rejectUnauthorized: true
};
this.tpmsClient = new cassandra.Client({
contactPoints: ['cassandra.eu-west-1.amazonaws.com'],
localDataCenter: 'eu-west-1',
authProvider: auth,
sslOptions: sslOptions1,
keyspace: 'tpms',
protocolOptions: { port: 9142 }
});
}
getOrganisation = async (orgKey) => {
const SQL = 'select * FROM organisation where organisation_id=?;';
return new Promise((resolve, reject) => {
this.tpmsClient.execute(SQL, [orgKey], {prepare: true}, (err, result) => {
if (!err?.message) resolve(result.rows);
else reject(err.message);
});
});
};
}
I am basically following this recommended AWS documentation.
https://docs.aws.amazon.com/keyspaces/latest/devguide/using_nodejs_driver.html
It seems that around 10-20% of the time the lambda function (cassandra driver) cannot connect to the endpoint.
I am pretty familiar with Cassandra (I already use a 6 node cluster that I manage) and don't have any issues with that.
Could this be a timeout or do I need more contact points?
Followed the recommended guides. Checked from the AWS console for any errors but none shown.
UPDATE:
Update to the above question....
I am occasionally (1 in 50 if I parallel call the function (5 concurrent calls)) getting the below error:
"All host(s) tried for query failed. First host tried,
3.248.244.5:9142: DriverError: Socket was closed at Connection.clearAndInvokePending
(/opt/node_modules/cassandra-driver/lib/connection.js:265:15) at
Connection.close
(/opt/node_modules/cassandra-driver/lib/connection.js:618:8) at
TLSSocket.
(/opt/node_modules/cassandra-driver/lib/connection.js:93:10) at
TLSSocket.emit (node:events:525:35)\n at node:net:313:12\n at
TCP.done (node:_tls_wrap:587:7) { info: 'Cassandra Driver Error',
isSocketError: true, coordinator: '3.248.244.5:9142'}
This exception may be caused by throttling in the keyspaces side, resulting the Driver Error that you are seeing sporadically.
I would suggest taking a look over this repo which should help you to put measures in place to either prevent the occurrence of this issue or at least reveal the true cause of the exception.
Some of the errors you see in the logs you will need to investigate Amazon CloudWatch metrics to see if you have throttling or system errors. I've built this AWS CloudFormation template to deploy a CloudWatch dashboard with all the appropriate metrics. This will provide better observability for your application.
A System Error indicates an event that must be resolved by AWS and often part of normal operations. Activities such as timeouts, server faults, or scaling activity could result in server errors. A User error indicates an event that can often be resolved by the user such as invalid query or exceeding a capacity quota. Amazon Keyspaces passes the System Error back as a Cassandra ServerError. In most cases this a transient error, in which case you can retry your request until it succeeds. Using the Cassandra driver’s default retry policy customers can also experience NoHostAvailableException or AllNodesFailedException or messages like yours "All host(s) tried for query failed". This is a client side exception that is thrown once all host in the load balancing policy’s query plan have attempted the request.
Take a look at this retry policy for NodeJs which should help resolve your "All hosts failed" exception or pass back the original exception.
The retry policies in the Cassandra drivers are pretty crude and will not be able to do more sophisticated things like circuit breaker patters. You may want to eventually use a "failfast" retry policy for the driver and handle the exceptions in your application code.

How to retrieve current workers count for job in GCP dataflow using API

Does anyone know if there is a possibility to get current workers count for active job that is running in GCP Dataflow?
I wasn't able to do it using provided by google API.
One thing that I was able to get is CurrentVcpuCount but it is not what I need.
Thanks in advance!
The current number of workers in a Dataflow job are displayed in the message logs, under autoscaling. For example, I did a quick job as example and I got the following message, when displaying the job logs in my Cloud Shell:
INFO:root:2019-01-28T16:42:33.173Z: JOB_MESSAGE_DETAILED: Autoscaling: Raised the number of workers to 0 based on the rate of progress in the currently running step(s).
INFO:root:2019-01-28T16:43:02.166Z: JOB_MESSAGE_DETAILED: Autoscaling: Raised the number of workers to 1 based on the rate of progress in the currently running step(s).
INFO:root:2019-01-28T16:43:05.385Z: JOB_MESSAGE_DETAILED: Workers have started successfully.
INFO:root:2019-01-28T16:43:05.433Z: JOB_MESSAGE_DETAILED: Workers have started successfully.
Now, you can query these messages by using the projects.jobs.messages.list method, in the Data flow API, and setting the minimumImportance parameter to be JOB_MESSAGE_BASIC.
You will get a response similar to the following:
...
"autoscalingEvents": [
{...} //other events
{
"currentNumWorkers": "1",
"eventType": "CURRENT_NUM_WORKERS_CHANGED",
"description": {
"messageText": "(fcfef6769cff802b): Worker pool started.",
"messageKey": "POOL_STARTUP_COMPLETED"
},
"time": "2019-01-28T16:43:02.130129051Z",
"workerPool": "Regular"
},
To extend this you could create a python script to parse the response, and only get the parameter currentNumWorkers from the last element in the list autoscalingEvents, to know what is the last (hence the current) number of workers in the Job.
Note that if this parameter is not present, it means that the number of workers is zero.
Edit:
I did a quick python script that retrieves the current number of workers, from the message logs, using the API I mentioned above:
from google.oauth2 import service_account
import googleapiclient.discovery
credentials = service_account.Credentials.from_service_account_file(
filename='PATH-TO-SERVICE-ACCOUNT-KEY/key.json',
scopes=['https://www.googleapis.com/auth/cloud-platform'])
service = googleapiclient.discovery.build(
'dataflow', 'v1b3', credentials=credentials)
project_id="MY-PROJECT-ID"
job_id="DATAFLOW-JOB-ID"
messages=service.projects().jobs().messages().list(
projectId=project_id,
jobId=job_id
).execute()
try:
print("Current number of workers is "+messages['autoscalingEvents'][-1]['currentNumWorkers'])
except:
print("Current number of workers is 0")
A couple of notes:
The scopes are the permissions needed on the service account key you are referencing (in the from_service_account_file function), in order to do the call to the API. This line is needed to authenticate to the API. You can use any one of this list, to make it easy on my side, I just used a service account key with project/owner permissions.
If you want to read more about the Python API Client Libraries, check this documentation, and this samples.
<script async src="//pagead2.googlesyndication.com/pagead/js/adsbygoogle.js"></script>
<script>
(adsbygoogle = window.adsbygoogle || []).push({
google_ad_client: "ca-pub-5513132861824326",
enable_page_level_ads: true
});
</script>

Unable to launch task from a spring cloud data flow stream

I registered my task app in Spring Cloud Data Flow, created a definition for it and the status shows 'unknown'. I created the stream and trying to launch the task through task-sink and I get an error:
java.lang.IllegalStateException: failed to resolve MavenResource:
How to launch a task from the task-sink? Am I missing something? Any help is appreciated. Another question I have is how do I access the payload sent via TaskLaunchRequest in my task?
S1 http | step1: transformer-rabbit | log
S2 :S1.step1 > filter --expression=payload.contains('CUSTADDRMODRQ_V15') | task-processor | task-sink
task-sink is launching the task provided by the uri in the TaskLaunchRequest. It is looking for the resource as shown in the log
OUT Using manager EnhancedLocalRepositoryManager with priority 10.0 for /home/vcap/.m2/repository
OUT Using transporter HttpTransporter with priority 5.0 for https://repo.spring.io/libs-snapshot and finally failing.
The task is deployed in our repository and as mentioned I registered and created the definition for it as well.
This one is in cf environment and I am using SCDF server 1.0.0.M4.
In the application.properties for the task-sink i am providing maven.remote.repositories.snapshots.url=**
task create fis-ifx-event-task --definition "fis-event-task"
My goal is launching the task from the stream.
Thanks for the information. I am in fact using the BUILD-SNAPSHOT as I am unable to enable taks in 1.0.0M4 version. Here is the one I am using spring-cloud-dataflow-server-cloudfoundry-1.0.0.BUILD-20160808.144306-116. I am able to register and create task definitions. The status of the task definition is showing as 'unknown' even when I am using the sample task module provided by your team. But when I initiate the flow of the stream and when task-sink tries to launch the task, it is unable to find the maven resource. When I create the task definition, does the task module gets deployed? I don't see any app in Pivotal Apps Manager. As mentioned earlier, I provided maven.remote.repositories.snapshot.url in the application.properties file for the task-sink application. Another thing I observed is when I launch the task manually from dataflow shell it gives an error CF-UnprocessableEntity(10008): The request is semantically invalid: Unknown field(s): 'staging_disk_in_mb', 'staging_memory_in_mb' and also a message saying 'Source is empty'. Presently the task is supposed to print the timestamp and is not dependent on any input.
TaskProcessor code:
#EnableBinding(Processor.class)
#EnableConfigurationProperties(TaskProcessorProperties.class)
public class TaskProcessor {
#Autowired
private TaskProcessorProperties processorProperties;
public TaskProcessor() {
}
#Transformer(inputChannel = Processor.INPUT, outputChannel = Processor.OUTPUT)
#ELI(level = "info", eventType = ELIEventType.INBOUND)
public Object setupRequest(String message) {
Map<String, String> properties = new HashMap<String, String>();
properties.put("payload", message);
TaskLaunchRequest request = new TaskLaunchRequest(processorProperties.getUri(), null, properties, null);
return new GenericMessage<>(request);
}
}
TaskSink code:
#SpringBootApplication
#EnableTaskLauncher
#EnableBinding(Sink.class)
#EnableConfigurationProperties(TaskSinkProperties.class)
public class FisIfxEventTaskSinkApplication {
public static void main(String[] args) {
SpringApplication.run(FisIfxEventTaskSinkApplication.class, args);
}
}
I provided the stream I am using earlier in the post. Sink is receiving the TaskLaunchRequest with uri and payload as you can see here and unable to launch the task.
OUT registering [40, java.io.File] with serializer org.springframework.integration.codec.kryo.FileSerializer
2016-08-10T16:08:55.02-0600 [APP/0]
OUT Launching Task for the following resource TaskLaunchRequest{uri='maven://com.xxx:fis.ifx.event-task:jar:1.0-SNAPSHOT', commandlineArguments=[], environmentProperties={payload={"statusCode":0,"fisT
opic":"CustomerDataUpdated","payloadId":"CUSTADDRMODR``Q_V15","customerIds":[1597304]}}, deploymentProperties={}}
Before I begin, you have a number of questions here. In the future, it's better to break them up into multiple questions so that they are easier to find by other users and easier to answer. That being said:
A little context on the current state of things
In order to understand how things will work, it's important to understand the current state of things. The current releases of the software involved are:
Pivotal Cloud Foundry (PCF) - 1.7.12. This version is required for any task support.
Spring Cloud Task (SCT) - 1.0.2.RELEASE
Spring Cloud Data Flow CF (SCDF) - 1.0.0.BUILD-SNAPSHOT (current as of the date of this post).
Currently PCF 1.7.12+ has all the capabilities to run tasks. You can create v3 applications (the type of application used to launch a task), run it as a task, etc. However, the tooling around that functionality is not currently complete. There is no support for v3 applications in Apps Manager or the CLI. There is a plugin for the CLI that is more of a dev tool that can be used to help with some functions (it will show you logs, etc), but it is not fully functional and requires a specific version of the CLI to work [1]. This is one of the reasons that the task functionality within PCF is still considered experimental.
Spring Cloud Task is currently GA and supports all the functionality needed to effectively run tasks on CF. However, it's important to note that SCT doesn't handle orchestration so the actual launching of tasks on CF is the responsibility of either the user, or Spring Cloud Data Flow (the easier route).
Spring Cloud Data Flow's Cloud Foundry server implementation currently has functionality to launch tasks on PCF in the latest snapshots. We have validated this against 1.7.12 as well as the development branch of 1.8.
The task workflow within SCDF
Tasks are fundamentally different from stream applications within the context of SCDF. When you create a stream definition, you are given the option to deploy it. What this does is it actually downloads the Spring Boot über jars and deploys them to PCF as long running processes. If they go down, PCF, will relaunch them as expected, etc.
Tasks on the other hand, are not deployed. They are launched. The difference is that while you create a task definition, there is nothing deployed until you click launch. And when the task completes, the software is shut down and cleaned up. So while a stream definition may have states, it's really a one to one relationship between the definition and the deployed software. Where with a task, you can launch a task definition as many times as you want.
Your issues
Reading through your post, I see a few things that you are struggling with. Let me see if I can help:
Task Definitions within SCDF and launching them via a stream - When launching a task from a stream, the task registry within SCDF is not used. The sink expects the URL for the resource to be within the TaskLauchRequest.
Apps Manager and tasks - As mentioned above, there is no support for v3 applications in Apps Manager yet so you won't be able to see your tasks there.
Viewing the logs - In order to debug what's going wrong with launching your task on CF, you're going to want to view the logs. To do so, use the v3 CLI plugin mentioned above to view them. It's important to note that you can only tail live logs with the plugin, not view logs that have previously been rendered. Because of that, when testing, you'll want to tail the logs as soon as the app is created, before it's launched.
Error in SCDF Shell - The error you received from the SCDF shell (CF-UnprocessableEntity(10008):...) leads me to wonder if you have both the correct version of PCF (1.7.12+) and the correct version of the following other libraries:
spring-cloud-deployer-cloudfoundry - The latest snapshots
cf-java-client - 2.0.0.M10+
reactor-core - 3.0.0.RC1+
I hope this helps!
[1] https://github.com/cloudfoundry/v3-cli-plugin
Task support is not available in 1.0.0.M4 release of SCDF's CF-server. In this release, the task commands/REST-APIs should be disabled - see here. And for that reason, you wouldn't see any docs related to Tasks in the 1.0.0.M4 reference guide.
That said, the Task support is available/enabled in the BUILD-SNAPSHOT release. If you're locally building the CF-server and upon pushing it to CF, you could take advantage the task commands in the shell to create and launch task definitions.

gae long time for calculation mapreduce 500 error python

I got a code to work on GAE but am struggling with the 500 error, which looks like due to the long wait (run) time.
I am doing the following:
Read the user given info
Run some mapreduce method to calculate some stats and send this as email
(Re)direct the user to a thank you page, since the results will be emailed
The code works fine on App engine SDK since there is no time limit. However, I keep getting the 500 error when I run the code on GAE. If I do not perform calculations in step 2 then the code works again (redirects to a new page and sends email). I tried doing step 2 after step 3, but keep getting the same error.
Is there any easy way to fix this? I am thinking of something like get the user info and let them know the results will be emailed to them or redirect them to the main page. In the meantime (or after the above) I can run mapreduce in the backend and email the completed results so the time limit does not abort my code.
class Guestbook(webapp2.RequestHandler):
def post(self):
#get info provided in form by user (code not shown here)
# send them to new page or main page
self.response.write('<html><body>You wrote:<pre>')
self.response.write("thanks")
self.response.write('</pre></body></html>')
#self.redirect('/')
dump_content = 'Error'
try:
dump_content = long_time_taking_mapreduce_method(user_given_info)
except DeadlineExceededError:
logging.warning("Deadline error")
send_results_as_email(OUTPFILE, dump_content)
app = webapp2.WSGIApplication([
('/', MainPage),
('/sign', Guestbook),
], debug=True)
The whole point of mapreduce is that it runs offline, taking as many tasks and as long as necessary. It's defeating the whole purpose to try and run it within your handler function.
Instead, your mapreduce task itself should call the send_results_as_email method once it has a result.

Triggering Celery Tasks from c++ with Redis

We have a setup where we have a web frontend programmed in Django and a backend written in C++ that parses data for us.
The frontend uses Celery in combination with Redis for asynchronous tasks.
Since it would be convenient in some situations, I was wondering today if it is possible to trigger a Celery task from within C++.
Since there is a Redis client available for C++, I am pretty sure that this is possible, if the correct messages are sent to Redis, however, I was not able to find any information on this anywhere.
My next step would be to try and dig the needed Information out of the Celery source code, but before I do that:
Does anybody have any information on this subject that could help me or get me started or is there even someone who has done this before?
Any help is appreciated. (Also if you got a reason why this will not work.)
Thank you.
I had a similar need to trigger a celery task from logstash. Basically, I had to create a message that looked something like this:
{
"body": "base_64_encoded_string (see below)",
"content-type": "application/json",
"properties": {
"body_encoding": "base64",
"correlation_id":"f009c9e0-0ca6-42a6-a046-3d0e53e06060",
"reply_to":"e1eb91f0-6780-4c34-b633-7ef9a46baf5e",
"delivery_mode":2,
"delivery_tag": "7788b924-a7fe-4c9a-839e-1c7ca602dbba",
"delivery_info": {
"priority":0,
"routing_key":"default",
"exchange":"default"
}
}
}
In this case, the decoded body translates to:
{
"args": ["meta_val","doc_value"],
"task":"goldstone.compliance.tasks.process_fim_event",
"id":"23deb69e-49c1-4a61-8639-d4627d0fc591"
}
If you have kwargs on your task, you can add a kwargs: {"key": "value", ...} to your body.
The body above maps triggers a task called process_fim_event. The task def looks like:
#task()
def process_fim_event(meta, doc):
...
The easiest way of doing this that I know of is to use flower, a HTTP Celery API. With flower you can create a task with anything that can make an HTTP request. One example from the Github Readme:
$ curl -X POST -d '{"args":[1,2]}' http://localhost:5555/api/task/async-apply/tasks.add
So, the idea is that your c++ app would make an HTTP request against the flower api, which would then insert the task into your Redis queue.