BigQuery results run via python in Google Cloud don't match results running on MAC - python-2.7

Question

I have a python app that runs a query on BigQuery and appends results to a file. I've run this on MAC workstation (Yosemite) and on GC instance (ubuntu 14.1) and the results for floating point differ. How can I make them the same? They python environments are the same on both.
run on google cloud instance
1120224,2015-04-06,23989,866,55159.71274162368,0.04923989554019882,0.021414467106578683,0.03609987911125933,63.69481840834143
54897577,2015-04-06,1188089,43462,2802473.708558333,0.051049132980100984,0.021641920553251377,0.03658143455582873,64.4810111950286
run on mac workstation
1120224,2015-04-06,23989,866,55159.712741623654,0.049239895540198794,0.021414467106578683,0.03609987911125933,63.694818408341405
54897577,2015-04-06,1188089,43462,2802473.708558335,0.05104913298010102,0.021641920553251377,0.03658143455582873,64.48101119502864
import sys
import pdb
import json
from collections import OrderedDict
from csv import DictWriter
from pprint import pprint
from apiclient import discovery
from oauth2client import tools
import functools
import argparse
import httplib2
import time
from subprocess import call
def authenticate_SERVICE_ACCOUNT(service_acct_email, private_key_path):
""" Generic authentication through a service accounts.
Args:
service_acct_email: The service account email associated
with the private key private_key_path: The path to the private key file
"""
from oauth2client.client import SignedJwtAssertionCredentials
with open(private_key_path, 'rb') as pk_file:
key = pk_file.read()
credentials = SignedJwtAssertionCredentials(
service_acct_email,
key,
scope='https://www.googleapis.com/auth/bigquery')
http = httplib2.Http()
auth_http = credentials.authorize(http)
return discovery.build('bigquery', 'v2', http=auth_http)
def create_query(number_of_days_ago):
""" Create a query
Args:
number_of_days_ago: Default value of 1 gets yesterday's data
"""
q = 'SELECT xxxxxxxxxx'
return q;
def translate_row(row, schema):
"""Apply the given schema to the given BigQuery data row.
Args:
row: A single BigQuery row to transform.
schema: The BigQuery table schema to apply to the row, specifically
the list of field dicts.
Returns:
Dict containing keys that match the schema and values that match
the row.
Adpated from bigquery client
https://github.com/tylertreat/BigQuery-Python/blob/master/bigquery/client.py
"""
log = {}
#pdb.set_trace()
# Match each schema column with its associated row value
for index, col_dict in enumerate(schema):
col_name = col_dict['name']
row_value = row['f'][index]['v']
if row_value is None:
log[col_name] = None
continue
# Cast the value for some types
if col_dict['type'] == 'INTEGER':
row_value = int(row_value)
elif col_dict['type'] == 'FLOAT':
row_value = float(row_value)
elif col_dict['type'] == 'BOOLEAN':
row_value = row_value in ('True', 'true', 'TRUE')
log[col_name] = row_value
return log
def extractResult(queryReply):
""" Extract a result from the query reply. Uses schema and rows to translate.
Args:
queryReply: the object returned by bigquery
"""
#pdb.set_trace()
result = []
schema = queryReply.get('schema', {'fields': None})['fields']
rows = queryReply.get('rows',[])
for row in rows:
result.append(translate_row(row, schema))
return result
def writeToCsv(results, filename, ordered_fieldnames, withHeader=True):
""" Create a csv file from a list of rows.
Args:
results: list of rows of data (first row is assumed to be a header)
order_fieldnames: a dict with names of fields in order desired - names must exist in results header
withHeader: a boolen to indicate whether to write out header -
Set to false if you are going to append data to existing csv
"""
try:
the_file = open(filename, "w")
writer = DictWriter(the_file, fieldnames=ordered_fieldnames)
if withHeader:
writer.writeheader()
writer.writerows(results)
the_file.close()
except:
print "Unexpected error:", sys.exc_info()[0]
raise
def runSyncQuery (client, projectId, query, timeout=0):
results = []
try:
print 'timeout:%d' % timeout
jobCollection = client.jobs()
queryData = {'query':query,
'timeoutMs':timeout}
queryReply = jobCollection.query(projectId=projectId,
body=queryData).execute()
jobReference=queryReply['jobReference']
# Timeout exceeded: keep polling until the job is complete.
while(not queryReply['jobComplete']):
print 'Job not yet complete...'
queryReply = jobCollection.getQueryResults(
projectId=jobReference['projectId'],
jobId=jobReference['jobId'],
timeoutMs=timeout).execute()
# If the result has rows, print the rows in the reply.
if('rows' in queryReply):
#print 'has a rows attribute'
#pdb.set_trace();
result = extractResult(queryReply)
results.extend(result)
currentPageRowCount = len(queryReply['rows'])
# Loop through each page of data
while('rows' in queryReply and currentPageRowCount < int(queryReply['totalRows'])):
queryReply = jobCollection.getQueryResults(
projectId=jobReference['projectId'],
jobId=jobReference['jobId'],
startIndex=currentRow).execute()
if('rows' in queryReply):
result = extractResult(queryReply)
results.extend(result)
currentRow += len(queryReply['rows'])
except AccessTokenRefreshError:
print ("The credentials have been revoked or expired, please re-run"
"the application to re-authorize")
except HttpError as err:
print 'Error in runSyncQuery:', pprint.pprint(err.content)
except Exception as err:
print 'Undefined error' % err
return results;
# Main
if __name__ == '__main__':
# Name of file
FILE_NAME = "results.csv"
# Default prior number of days to run query
NUMBER_OF_DAYS = "1"
# BigQuery project id as listed in the Google Developers Console.
PROJECT_ID = 'xxxxxx'
# Service account email address as listed in the Google Developers Console.
SERVICE_ACCOUNT = 'xxxxxx#developer.gserviceaccount.com'
KEY = "/usr/local/xxxxxxxx"
query = create_query(NUMBER_OF_DAYS)
# Authenticate
client = authenticate_SERVICE_ACCOUNT(SERVICE_ACCOUNT, KEY)
# Get query results
results = runSyncQuery (client, PROJECT_ID, query, timeout=0)
#pdb.set_trace();
# Write results to csv without header
ordered_fieldnames = OrderedDict([('f_split',None),('m_members',None),('f_day',None),('visitors',None),('purchasers',None),('demand',None), ('dmd_per_mem',None),('visitors_per_mem',None),('purchasers_per_visitor',None),('dmd_per_purchaser',None)])
writeToCsv(results, FILE_NAME, ordered_fieldnames, False)
# Backup current data
backupfilename = "data_bk-" + time.strftime("%y-%m-%d") + ".csv"
call(['cp','../data/data.csv',backupfilename])
# Concatenate new results to data
with open("../data/data.csv", "ab") as outfile:
with open("results.csv","rb") as infile:
line = infile.read()
outfile.write(line)

Answer 1

You mention that these come from aggregate sums of floating point data. As Felipe mentioned, floating point is awkward; it violates some of the mathematical identities that we tend to assume.
In this case, the associative property is the one that bites us. That is, usually (A+B)+C == A+(B+C). However, in floating point math, this isn't the case. Each operation is an approximation; you can see this better if you wrap with an 'approx' function: approx(approx(A+B) + C) is clearly different from approx(A + approx(B+C)).
If you think about how bigquery computes aggregates, it builds an execution tree, and computes the value to be aggregated at the leaves of the tree. As those answers are ready, they're passed back up to the higher levels of the tree and aggregated (let's say they're added). The "when they're ready" part makes it non-deterministic.
A node may get results back in the order A,B,C the first time and C,A, B the second time. This means that the order of distribution will change, since you'll get approx(approx(A + B) + C) the first time and approx(approx(C, A) + B) the second time. Note that since we're dealing with ordering, it may look like the commutative property is the problematic one, but it isn't; A+B in floating math is the same as B+A. The problem is really that you're adding partial results, which aren't associative.
Floating point math has all sorts of nasty properties and should usually be avoided if you rely on precision.

Answer 2

Assume floating point is non-deterministic:
https://randomascii.wordpress.com/2013/07/16/floating-point-determinism/
“the IEEE standard does not guarantee that the same program will
deliver identical results on all conforming systems.”