Resource Allocation for Incremental Pipelines - palantir-foundry

There are times when an incremental pipeline in Palantir Foundry has to be built as a snapshot. If the data size is large, the resources to run the build are increased to reduce run time and then the configuration is removed after first snapshot run. Is there a way to set conditional configuration? Like if pipeline is running on Incremental Mode, use default configuration of resource allocation and if not the specified set of resources.
Example:
If pipeline runs as snapshot transaction, below configuration has to be applied
#configure(profile=["NUM_EXECUTORS_8", "EXECUTOR_MEMORY_MEDIUM", "DRIVER_MEMORY_MEDIUM"])
If incremental, then the default one.

The #configure and #incremental are set during the CI execution, while the actual code inside the function annotated by #transform_df or `#transform happens at build time.
This means that you can't programatically switch between them after the CI has passed. What you can do however is have a constant or configuration within your repo, and switch at code level whenever you want to switch these. Please make sure you understand how semantic versioning works before attempting this I.e.:
IS_INCREMENTAL = true
SEMANTIC_VERSION=1
def mytransform(input1, input2,...)
return input1.join(input2, "foo", left)
if IS_INCREMENTAL:
#incremental(semantic_version=SEMANTIC_VERSION)
#transform_df(
Output("foo"),
input1=Input("bar"),
input2=Input("foobar"))
def compute(input1, input2):
return mytransform(input1, input2)
else:
#configure(profile=["NUM_EXECUTORS_8", "EXECUTOR_MEMORY_MEDIUM", "DRIVER_MEMORY_MEDIUM"])
#transform_df(
Output("foo"),
input1=Input("bar"),
input2=Input("foobar"))
def compute(input1, input2):
return mytransform(input1, input2)

Related

SQLAlchemy automap: Best practices for performance

I'm building a python app around an existing (mysql) database and am using automap to infer tables and relationships:
base = automap_base()
self.engine = create_engine(
'mysql://%s:%s#%s/%s?charset=utf8mb4' % (
config.DB_USER, config.DB_PASSWD, config.DB_HOST, config.DB_NAME
), echo=False
)
# reflect the tables
base.prepare(self.engine, reflect=True)
self.TableName = base.classes.table_name
Using this I can do things like session.query(TableName) etc...
However, I'm worried about performance, because every time the app runs it will do the whole inference again.
Is this a legitimate concern?
If so, is there a possibility to 'cache' the output of Automap?
I think that "reflecting" the structure of your database is not the way to go. Unless your app tries to "infer" things from the structure, like static code analysis would for source files, then it is unnecessary. The other reason for reflecting it at run-time would be the reduced time to begin "using" the database using SQLAlchemy. However:
Another option would be to use something like SQLACodegen (https://pypi.python.org/pypi/sqlacodegen):
It will "reflect" your database once and create a 99.5% accurate set of declarative SQLAlchemy models for you to work with. However, this does require that you keep the model subsequently in-sync with the structure of the database. I would assume that this is not a big concern seeing as the tables you're already-working with are stable-enough such that run-time reflection of their structure does not impact your program much.
Generating the declarative models is essentially a "cache" of the reflection. It's just that SQLACodegen saved it into a very readable set of classes + fields instead of data in-memory. Even with a changing structure, and my own "changes" to the generated declarative models, I still use SQLACodegen later-on in a project whenever I make database changes. It means that my models are generally consistent amongst one-another and that I don't have things such as typos and data-mismatches due to copy-pasting.
Performance can be a legitimate concern. If the database schema is not changing, it can be time consuming to reflect the database every time a script is run. This is more of an issue during development, not starting up a long running application. It's also a significant time saver if your database is on a remote server (again, particularly during development).
I use code that is similar to the answer here (as noted by #ACV). The general plan is to perform the reflection the first time, then pickle the metadata object. The next time the script is run, it will look for the pickle file and use that. The file can be anywhere, but I place mine in ~/.sqlalchemy_cache. This is an example based on your code.
import os
from sqlalchemy.ext.declarative import declarative_base
self.engine = create_engine(
'mysql://%s:%s#%s/%s?charset=utf8mb4' % (
config.DB_USER, config.DB_PASSWD, config.DB_HOST, config.DB_NAME
), echo=False
)
metadata_pickle_filename = "mydb_metadata"
cache_path = os.path.join(os.path.expanduser("~"), ".sqlalchemy_cache")
cached_metadata = None
if os.path.exists(cache_path):
try:
with open(os.path.join(cache_path, metadata_pickle_filename), 'rb') as cache_file:
cached_metadata = pickle.load(file=cache_file)
except IOError:
# cache file not found - no problem, reflect as usual
pass
if cached_metadata:
base = declarative_base(bind=self.engine, metadata=cached_metadata)
else:
base = automap_base()
base.prepare(self.engine, reflect=True) # reflect the tables
# save the metadata for future runs
try:
if not os.path.exists(cache_path):
os.makedirs(cache_path)
# make sure to open in binary mode - we're writing bytes, not str
with open(os.path.join(cache_path, metadata_pickle_filename), 'wb') as cache_file:
pickle.dump(Base.metadata, cache_file)
except:
# couldn't write the file for some reason
pass
self.TableName = base.classes.table_name
For anyone using declarative table class definitions, assuming a Base object defined as e.g.
Base = declarative_base(bind=engine)
metadata_pickle_filename = "ModelClasses_trilliandb_trillian.pickle"
# ------------------------------------------
# Load the cached metadata if it's available
# ------------------------------------------
# NOTE: delete the cached file if the database schema changes!!
cache_path = os.path.join(os.path.expanduser("~"), ".sqlalchemy_cache")
cached_metadata = None
if os.path.exists(cache_path):
try:
with open(os.path.join(cache_path, metadata_pickle_filename), 'rb') as cache_file:
cached_metadata = pickle.load(file=cache_file)
except IOError:
# cache file not found - no problem
pass
# ------------------------------------------
# define all tables
#
class MyTable(Base):
if cached_metadata:
__table__ = cached_metadata.tables['my_schema.my_table']
else:
__tablename__ = 'my_table'
__table_args__ = {'autoload':True, 'schema':'my_schema'}
...
# ----------------------------------------
# If no cached metadata was found, save it
# ----------------------------------------
if cached_metadata is None:
# cache the metadata for future loading
# - MUST DELETE IF THE DATABASE SCHEMA HAS CHANGED
try:
if not os.path.exists(cache_path):
os.makedirs(cache_path)
# make sure to open in binary mode - we're writing bytes, not str
with open(os.path.join(cache_path, metadata_pickle_filename), 'wb') as cache_file:
pickle.dump(Base.metadata, cache_file)
except:
# couldn't write the file for some reason
pass
Important Note!! If the database schema changes, you must delete the cached file to force the code to autoload and create a new cache. If you don't, the changes will be be reflected in the code. It's an easy thing to forget.
The answer to your first question is largely subjective. You are adding database queries to fetch the reflection metadata to the application load time. Whether or not that overhead is significant depends on your project requirements.
For reference, I have an internal tool at work that uses a reflection pattern because the the load-time is acceptable for our team. That might not be the case if it were an externally-facing product. My hunch is that for most applications the reflection overhead will not dominate the total application load time.
If you decide it is significant for your purposes, this question has an interesting answer where the user pickles the database metadata in order to locally cache it.
Adding to this, what #Demitri answered is close to correct but (at least in sqlalchemy 1.4.29), the example will fail on the last line self.TableName = base.classes.table_name when generating from the cached file. In this case declarative_base() has no attribute classes.
To fix is as simple as altering:
if cached_metadata:
base = declarative_base(bind=self.engine, metadata=cached_metadata)
else:
base = automap_base()
base.prepare(self.engine, reflect=True) # reflect the tables
to
if cached_metadata:
base = automap_base(declarative_base(bind=self.engine, metadata=cached_metadata))
base.prepare()
else:
base = automap_base()
base.prepare(self.engine, reflect=True) # reflect the tables
This will create your automap object with the proper attributes.

GCP dataflow - processing JSON takes too long

I am trying to process json files in a bucket and write the results into a bucket:
DataflowPipelineOptions options = PipelineOptionsFactory.create()
.as(DataflowPipelineOptions.class);
options.setRunner(BlockingDataflowPipelineRunner.class);
options.setProject("the-project");
options.setStagingLocation("gs://some-bucket/temp/");
Pipeline p = Pipeline.create(options);
p.apply(TextIO.Read.from("gs://some-bucket/2016/04/28/*/*.json"))
.apply(ParDo.named("SanitizeJson").of(new DoFn<String, String>() {
#Override
public void processElement(ProcessContext c) {
try {
JsonFactory factory = JacksonFactory.getDefaultInstance();
String json = c.element();
SomeClass e = factory.fromString(json, SomeClass.class);
// manipulate the object a bit...
c.output(factory.toString(e));
} catch (Exception err) {
LOG.error("Failed to process element: " + c.element(), err);
}
}
}))
.apply(TextIO.Write.to("gs://some-bucket/output/"));
p.run();
I have around 50,000 files under the path gs://some-bucket/2016/04/28/ (in sub-directories).
My question is: does it make sense that this takes more than an hour to complete? Doing something similar on a Spark cluster in amazon takes about 15-20 minutes. I suspect that I might be doing something inefficiently.
EDIT:
In my Spark job I aggregate all the results in a DataFrame and only then write the output, all at once. I noticed that my pipeline here writes each file separately, I assume that is why it's taking much longer. Is there a way to change this behavior?
Your jobs are hitting a couple of performance issues in Dataflow, caused by the fact that it is more optimized for executing work in larger increments, while your job is processing lots of very small files. As a result, some aspects of the job's execution end up dominated by per-file overhead. Here's some details and suggestions.
The job is limited rather by writing output than by reading input (though reading input is also a significant part). You can significantly cut that overhead by specifying withNumShards on your TextIO.Write, depending on how many files you want in the output. E.g. 100 could be a reasonable value. By default you're getting an unspecified number of files which in this case, given current behavior of the Dataflow optimizer, matches number of input files: usually it is a good idea because it allows us to not materialize the intermediate data, but in this case it's not a good idea because the input files are so small and per-file overhead is more important.
I recommend to set maxNumWorkers to a value like e.g. 12 - currently the second job is autoscaling to an excessively large number of workers. This is caused by Dataflow's autoscaling currently being geared toward jobs that process data in larger increments - it currently doesn't take into account per-file overhead and behaves not so well in your case.
The second job is also hitting a bug because of which it fails to finalize the written output. We're investigating, however setting maxNumWorkers should also make it complete successfully.
To put it shortly:
set maxNumWorkers=12
set TextIO.Write.to("...").withNumShards(100)
and it should run much better.

How to commit model instances and remove them from working memory a few at a time

I have a pyramid view that is used for loading data from a large file into a database. For each line in the file it does a little processing then creates some model instances and adds them to the session. This works fine except when the files are big. For large files the view slowly eats up all my ram until everything effectively grinds to a halt.
So my idea is to process each line individually with a function that creates a session, creates the necessary model instances and adds them to the current session, then commits.
def commit_line(lTitles,lLine,oStartDate,oEndDate,iDS,dSettings):
from sqlalchemy.orm import (
scoped_session,
sessionmaker,
)
from sqlalchemy import engine_from_config
from pyramidapp.models import Base, DataEntry
from zope.sqlalchemy import ZopeTransactionExtension
import transaction
oCurrentDBSession = scoped_session(sessionmaker(extension=ZopeTransactionExtension()))
engine = engine_from_config(dSettings, 'sqlalchemy.')
oCurrentDBSession.configure(bind=engine)
Base.metadata.bind = engine
oEntry = DataEntry()
oCurrentDBSession.add(oEntry)
...
transaction.commit()
My requirements for this function are as follows:
create a session (check)
make a bunch of model instances (check)
add those instances to the session (check)
commit those models to the database
get rid of the session (so that it and the objects created in 2 are garbage collected)
I've made sure that the newly created session is passed as an argument whenever necessary in order to stop errors to do with multiple sessions blah blah. But alas! I can't get database connections to go away and stuff isn't being committed.
I tried separating the function out into a celery task so the view executes to completion and does what it needs to but I'm getting an error in celery about having too many mysql connections no matter what I try in terms of committing and closing and disposing and I'm not sure why. And yes, I restart the celery server when I make changes.
Surely there is a simple way to do this? All I want to do is make a session commit then go away and leave me alone.
Creating a new session for each line of your large file is going to be quite slow I would imagine.
What I would try is to commit the session and expunge all objects from it every 1000 rows or so:
counter = 0
for line in mymegafile:
entry = process_line(line)
session.add(entry)
if counter > 1000:
counter = 0
transaction.commit() # if you insist on using ZopeTransactionExtension, otherwise session.commit()
session.expunge_all() # this may not be required actually, see https://groups.google.com/forum/#!topic/sqlalchemy/We4XGX2CYX8
else:
counter += 1
If there are no references to DataEntry instances from anywhere they should be garbage collected by Python interpreter at some point.
However, if all you're doing in that view is inserting new records to the database, it may be much more efficient to use SQLAlchemy Core constructs or literal SQL to bulk-insert data. This would also get rid of the problem with your ORM instances eating up your RAM. See I’m inserting 400,000 rows with the ORM and it’s really slow! for details.
So I tried a bunch of things and, although using SQLAlchemy's built in functionality to solve this was probably possible I could not find any way of pulling that off.
So here's an outline of what I did:
seperate the lines to be processed into batches
for each batch of lines queue up a celery task to deal with those lines
in the celery task a seperate process is launched that does the necessary stuff with the lines.
Reasoning:
The batch stuff is obvious
Celery was used because it took a heck of a long time to process an entire file so queuing just made sense
the task launched a separate process because if it didn't then I had the same problem that I had with the pyramid application
Some code:
Celery task:
def commit_lines(lLineData,dSettings,cwd):
"""
writes the line data to a file then calls a process that reads the file and creates
the necessary data entries. Then deletes the file
"""
import lockfile
sFileName = "/home/sheena/tmp/cid_line_buffer"
lock = lockfile.FileLock("{0}_lock".format(sFileName))
with lock:
f = open(sFileName,'a') #in case the process was at any point interrupted...
for d in lLineData:
f.write('{0}\n'.format(d))
f.close()
#now call the external process
import subprocess
import os
sConnectionString = dSettings.get('sqlalchemy.url')
lArgs = [
'python',os.path.join(cwd,'commit_line_file.py'),
'-c',sConnectionString,
'-f',sFileName
]
#open the subprocess. wait for it to complete before continuing with stuff. if errors: raise
subprocess.check_call(lArgs,shell=False)
#and clear the file
lock = lockfile.FileLock("{0}_lock".format(sFileName))
with lock:
f = open(sFileName,'w')
f.close()
External process:
"""
this script goes through all lines in a file and creates data entries from the lines
"""
def main():
from optparse import OptionParser
from sqlalchemy import create_engine
from pyramidapp.models import Base,DBSession
import ast
import transaction
#get options
oParser = OptionParser()
oParser.add_option('-c','--connection_string',dest='connection_string')
oParser.add_option('-f','--input_file',dest='input_file')
(oOptions, lArgs) = oParser.parse_args()
#set up connection
#engine = engine_from_config(dSettings, 'sqlalchemy.')
engine = create_engine(
oOptions.connection_string,
echo=False)
DBSession.configure(bind=engine)
Base.metadata.bind = engine
#commit stuffs
import lockfile
lock = lockfile.FileLock("{0}_lock".format(oOptions.input_file))
with lock:
for sLine in open(oOptions.input_file,'r'):
dLine = ast.literal_eval(sLine)
create_entry(**dLine)
transaction.commit()
def create_entry(iDS,oStartDate,oEndDate,lTitles,lValues):
#import stuff
oEntry = DataEntry()
#do some other stuff, make more model instances...
DBSession.add(oEntry)
if __name__ == "__main__":
main()
in the view:
for line in big_giant_csv_file_handler:
lLineData.append({'stuff':'lots'})
if lLineData:
lLineSets = [lLineData[i:i+iBatchSize] for i in range(0,len(lLineData),iBatchSize)]
for l in lLineSets:
commit_lines.delay(l,dSettings,sCWD) #queue it for celery
You are just doing it wrong. Period.
Quoted from SQLAlchemy docs
The advanced developer will try to keep the details of session,
transaction and exception management as far as possible from the
details of the program doing its work.
Quoted from Pyramid docs
We made the decision to use SQLAlchemy to talk to our database. We also, though, installed pyramid_tm and zope.sqlalchemy.
Why?
Pyramid has a strong orientation towards support for transactions.
Specifically, you can install a transaction manager into your app
application, either as middleware or a Pyramid "tween". Then, just
before you return the response, all transaction-aware parts of your
application are executed. This means Pyramid view code usually doesn't
manage transactions.
My answer today is not code, but a recommendation to follow best practices recommended by the authors of the packages/frameworks you are working with.
References
Big picture - Using Thread-Local Scope with Web Applications
Typical error message when doing it wrong
Databases using SQLAlchemy
How to use scoped_session
Encapsulate CSV reading and creating SQLAlchemy model instances into something that supports the iterator protocol. I called it BatchingModelReader. It returns a collection of DataEntry instances, collection size depends on batch size. If the model changes overtime, you do not need to change the celery task. The task only puts a batch of models into a session and commits the transaction. By controlling the batch size you control memory consumption. Neither BatchingModelReader nor the celery task save huge amounts of intermediate data. This example shows as well that using celery is only an option. I added links to code samples of an pyramid application I am actually refactoring in a Github fork.
BatchingModelReader - encapsulates csv.reader and uses existing models from your pyramid application
get inspired by source code of csv.DictReader
could be run as a celery task - use appropriate task decorator
from .models import DBSession
import transaction
def import_from_csv(path_to_csv, batchsize)
"""given a CSV file and batchsize iterate over batches of model instances and import them to database"""
for batch in BatchingModelReader(path_to_csv, batchsize):
with transaction.manager:
DBSession.add_all(batch)
pyramid view - just save big giant CSV file, start task, return response immediately
#view_config(...):
def view(request):
"""gets file from request, save it to filesystem and start celery task"""
with open(path_to_csv, 'w') as f:
f.write(big_giant_csv_file)
#start task with parameters
import_from_csv.delay(path_to_csv, 1000)
Code samples
ToDoPyramid - commit transaction from commandline
ToDoPyramid - commit transaction from request
Pyramid using SQLAlchemy
Databases using SQLAlchemy
SQLAlchemy internals
Big picture - Using Thread-Local Scope with Web Applications
How to use scoped_session

Returning values from InputFormat via the Hadoop Configuration object

Consider a running Hadoop job, in which a custom InputFormat needs to communicate ("return", similarly to a callback) a few simple values to the driver class (i.e., to the class that has launched the job), from within its overriden getSplits() method, using the new mapreduce API (as opposed to mapred).
These values should ideally be returned in-memory (as opposed to saving them to HDFS or to the DistributedCache).
If these values were only numbers, one could be tempted to use Hadoop counters. However, in numerous tests counters do not seem to be available at the getSplits() phase and anyway they are restricted to numbers.
An alternative could be to use the Configuration object of the job, which, as the source code reveals, should be the same object in memory for both the getSplits() and the driver class.
In such a scenario, if the InputFormat wants to "return" a (say) positive long value to the driver class, the code would look something like:
// In the custom InputFormat.
public List<InputSplit> getSplits(JobContext job) throws IOException
{
...
long value = ... // A value >= 0
job.getConfiguration().setLong("value", value);
...
}
// In the Hadoop driver class.
Job job = ... // Get the job to be launched
...
job.submit(); // Start running the job
...
while (!job.isComplete())
{
...
if (job.getConfiguration().getLong("value", -1))
{
...
}
else
{
continue; // Wait for the value to be set by getSplits()
}
...
}
The above works in tests, but is it a "safe" way of communicating values?
Or is there a better approach for such in-memory "callbacks"?
UPDATE
The "in-memory callback" technique may not work in all Hadoop distributions, so, as mentioned above, a safer way is, instead of saving the values to be passed back in the Configuration object, create a custom object, serialize it (e.g., as JSON), saved it (in HDFS or in the distributed cache) and have it read in the driver class. I have also tested this approach and it works as expected.
Using the configuration is a perfectly suitable solution (admittedly for a problem I'm not sure I understand), but once the job has actually been submitted to the Job tracker, you will not be able to amend this value (client side or task side) and expect to see the change on the opposite side of the comms (setting configuration values in a map task for example will not be persisted to the other mappers, nor to the reducers, nor will be visible to the job tracker).
So to communicate information back from within getSplits back to your client polling loop (to see when the job has actually finished defining the input splits) is fine in your example.
What's your greater aim or use case for using this?

SQL Alchemy + Testing webserver with InnoDB fails

I am currently trying to move my DB tables over to InnoDB from MyISAM. I am having timing issues with requests and cron jobs that are running on the server that is leading to some errors. I am quite sure that transaction support will help me with the problem. I am therefore transitioning to InnoDB.
I have a suite of tests which make calls to our webservices REST API and receive XML responses. The test suite is fairly thorough, and it's written in Python and uses SQLAlchemy to query information from the database. When I change the tables in the system from MyISAM to InnoDB however, the tests start failing. However, the tests aren't failing because the system isn't working, they are failing because the ORM is not correctly querying the rows from the database I am testing on. when I step through the code I see the correct results, but the ORM is not returning the correct results at all.
Basic flow is:
class UnitTest(unittest.TestCase):
def setUp(self):
# Create a test object in DB that gets affected by the web server
testObject = Obj(foo='one')
self.testId = testObject.id
session.add(testObject)
session.commit()
def tearDown(self):
# Clean up after the test
testObject = session.query(Obj).get(self.testId)
session.delete(testObject)
session.commit()
def test_web_server(self):
# Ensure the initial state of the object.
objects = session.query(Obj).get(self.testId)
assert objects.foo == 'one'
# This will make a simple HTTP get call on an url that will modify the DB
response = server.request.increment_foo(self.testId)
# This one fails, the object still has a foo of 'one'
# When I stop here in a debugger though, and look at the database,
# The row in question actually has the correct value in the database.
# ????
objects = session.query(Obj).get(self.testId)
assert objects.foo == 'two'
Using MyISAM tables to store the object and this test will pass. However, when I change to InnoDB tables, this test will not pass. What is more interesting is that when I step through the code in the debugger, I can see that the datbase has what I expect, so it's not a problem in the web server code. I have tried nearly every combination of expire_all, autoflush, autocommit, etc. etc, and still can't get this test to pass.
I can provide more info if necessary.
Thanks,
Conrad
The problem is that you put the line self.testId = testObject.id before new object is added to session, flushed, and SQLAlchemy assigned ID to it. Thus self.testId is always None. Move this line below session.commit().