google dataflow read from spanner

Question

I am trying to read a table from a Google spanner database, and write it to a text file to do a backup, using google dataflow with the python sdk. I have written the following script:

    from __future__ import absolute_import

import argparse
import itertools
import logging
import re
import time
import datetime as dt
import logging

import apache_beam as beam
from apache_beam.io import iobase
from apache_beam.io import WriteToText
from apache_beam.io.range_trackers import OffsetRangeTracker, UnsplittableRangeTracker
from apache_beam.metrics import Metrics
from apache_beam.options.pipeline_options import PipelineOptions
from apache_beam.options.pipeline_options import StandardOptions, SetupOptions
from apache_beam.options.pipeline_options import GoogleCloudOptions

from google.cloud.spanner.client import Client
from google.cloud.spanner.keyset import KeySet

BUCKET_URL = 'gs://my_bucket'
OUTPUT = '%s/output/' % BUCKET_URL
PROJECT_ID = 'my_project'
INSTANCE_ID = 'my_instance'
DATABASE_ID = 'my_db'
JOB_NAME = 'spanner-backup'
TABLE = 'my_table'


class SpannerSource(iobase.BoundedSource):
    def __init__(self):
    logging.info('Enter __init__')

    self.spannerOptions = {
        "id": PROJECT_ID,
        "instance": INSTANCE_ID,
        "database": DATABASE_ID
    }
    self.SpannerClient = Client

    def estimate_size(self):
    logging.info('Enter estimate_size')
    return 1

    def get_range_tracker(self, start_position=None, stop_position=None):
    logging.info('Enter get_range_tracker')
    if start_position is None:
       start_position = 0
    if stop_position is None:
       stop_position = OffsetRangeTracker.OFFSET_INFINITY

    range_tracker = OffsetRangeTracker(start_position, stop_position)
    return UnsplittableRangeTracker(range_tracker)

    def read(self, range_tracker):  # This is not called when using the dataflowRunner !
    logging.info('Enter read')
    # instantiate spanner client
    spanner_client = self.SpannerClient(self.spannerOptions["id"])
    instance = spanner_client.instance(self.spannerOptions["instance"])
    database = instance.database(self.spannerOptions["database"])

    # read from table
    table_fields = database.execute_sql("SELECT t.column_name FROM information_schema.columns AS t WHERE t.table_name = '%s'" % TABLE)
    table_fields.consume_all()
    self.columns = [x[0] for x in table_fields]
    keyset = KeySet(all_=True)
    results = database.read(table=TABLE, columns=self.columns, keyset=keyset)

    # iterator over rows
    results.consume_all()
    for row in results:
        JSON_row = {
        self.columns[i]: row[i] for i in range(len(self.columns))
        }
        yield JSON_row

    def split(self, start_position=None, stop_position=None):
    # this should not be called since the source is unspittable
    logging.info('Enter split')
    if start_position is None:
        start_position = 0
    if stop_position is None:
        stop_position = 1

    # Because the source is unsplittable (for now), only a single source is returned
    yield iobase.SourceBundle(
        weight=1,
        source=self,
        start_position=start_position,
        stop_position=stop_position)


def run(argv=None):
  """Main entry point"""
  pipeline_options = PipelineOptions()
  google_cloud_options = pipeline_options.view_as(GoogleCloudOptions)
  google_cloud_options.project = PROJECT_ID
  google_cloud_options.job_name = JOB_NAME
  google_cloud_options.staging_location = '%s/staging' % BUCKET_URL
  google_cloud_options.temp_location = '%s/tmp' % BUCKET_URL

  #pipeline_options.view_as(StandardOptions).runner = 'DirectRunner'
  pipeline_options.view_as(StandardOptions).runner = 'DataflowRunner'
  p = beam.Pipeline(options=pipeline_options)

  output = p | 'Get Rows from Spanner' >> beam.io.Read(SpannerSource())
  iso_datetime = dt.datetime.now().replace(microsecond=0).isoformat()
  output | 'Store in GCS' >> WriteToText(file_path_prefix=OUTPUT + iso_datetime + '-' + TABLE, file_name_suffix='')  # if this line is commented, job completes but does not do anything


  result = p.run()
  result.wait_until_finish()


if __name__ == '__main__':
  logging.getLogger().setLevel(logging.INFO)
  run()

However, this script runs correctly only on the DirectRunner: when I let it run on the DataflowRunner, it runs for a while without any output, before exiting with an error:

"Executing failure step failure14 [...] Workflow failed. Causes: [...] The worker lost contact with the service."

Sometimes, it just goes on forever, without creating an output.

Moreover, if I comment the line 'output = ...', the job completes, but without actually reading the data.

It also appears that the dataflowRunner calls the function 'estimate_size' of the source, but not the functions 'read' or 'get_range_tracker'.

Does anyone have any ideas about what may cause this ? I know there is a (more complete) java SDK with an experimental spanner source/sink available, but if possible I'd rather stick with python.

Thanks

Answer 1

Google currently added support of Backup Spanner with Dataflow, you can choose related template when creating DataFlow job.

For more: https://cloud.google.com/blog/products/gcp/cloud-spanner-adds-import-export-functionality-to-ease-data-movement

Answer 2

I have reworked my code following the suggestion to simply use a ParDo, instead of using the BoundedSource class. As a reference, here is my solution; I am sure there are many ways to improve on it, and I would be happy to to hear opinions. In particular I am surprised that I have to a create a dummy PColl when starting the pipeline (if I don't, I get an error

AttributeError: 'PBegin' object has no attribute 'windowing'

that I could not work around. The dummy PColl feels a bit like a hack.

from __future__ import absolute_import

import datetime as dt
import logging

import apache_beam as beam
from apache_beam.io import WriteToText
from apache_beam.options.pipeline_options import PipelineOptions
from apache_beam.options.pipeline_options import StandardOptions, SetupOptions
from apache_beam.options.pipeline_options import GoogleCloudOptions
from google.cloud.spanner.client import Client
from google.cloud.spanner.keyset import KeySet

BUCKET_URL = 'gs://my_bucket'
OUTPUT = '%s/some_folder/' % BUCKET_URL
PROJECT_ID = 'my_project'
INSTANCE_ID = 'my_instance'
DATABASE_ID = 'my_database'
JOB_NAME = 'my_jobname'

class ReadTables(beam.DoFn):
    def __init__(self, project, instance, database):
        super(ReadTables, self).__init__()
        self._project = project
        self._instance = instance
        self._database = database

    def process(self, element):
        # get list of tables in the database
        table_names_row = Client(self._project).instance(self._instance).database(self._database).execute_sql('SELECT t.table_name FROM information_schema.tables AS t')
        for row in table_names_row:
            if row[0] in [u'COLUMNS', u'INDEXES', u'INDEX_COLUMNS', u'SCHEMATA', u'TABLES']:    # skip these
                continue
            yield row[0]

class ReadSpannerTable(beam.DoFn):
    def __init__(self, project, instance, database):
        super(ReadSpannerTable, self).__init__()
        self._project = project
        self._instance = instance
        self._database = database

    def process(self, element):
        # first read the columns present in the table
        table_fields = Client(self._project).instance(self._instance).database(self._database).execute_sql("SELECT t.column_name FROM information_schema.columns AS t WHERE t.table_name = '%s'" % element)
        columns = [x[0] for x in table_fields]

        # next, read the actual data in the table
        keyset = KeySet(all_=True)
        results_streamed_set = Client(self._project).instance(self._instance).database(self._database).read(table=element, columns=columns, keyset=keyset)

        for row in results_streamed_set:
            JSON_row = { columns[i]: row[i] for i in xrange(len(columns)) }
            yield (element, JSON_row)            # output pairs of (table_name, data)

def run(argv=None):
  """Main entry point"""
  pipeline_options = PipelineOptions()
  pipeline_options.view_as(SetupOptions).save_main_session = True
  pipeline_options.view_as(SetupOptions).requirements_file = "requirements.txt"
  google_cloud_options = pipeline_options.view_as(GoogleCloudOptions)
  google_cloud_options.project = PROJECT
  google_cloud_options.job_name = JOB_NAME
  google_cloud_options.staging_location = '%s/staging' % BUCKET_URL
  google_cloud_options.temp_location = '%s/tmp' % BUCKET_URL

  pipeline_options.view_as(StandardOptions).runner = 'DataflowRunner'
  p = beam.Pipeline(options=pipeline_options)

  init   = p       | 'Begin pipeline'              >> beam.Create(["test"])                                                 # have to create a dummy transform to initialize the pipeline, surely there is a better way ?
  tables = init    | 'Get tables from Spanner'     >> beam.ParDo(ReadTables(PROJECT, INSTANCE_ID, DATABASE_ID))          # read the tables in the db
  rows = (tables   | 'Get rows from Spanner table' >> beam.ParDo(ReadSpannerTable(PROJECT, INSTANCE_ID, DATABASE_ID))    # for each table, read the entries
                   | 'Group by table'              >> beam.GroupByKey()
                   | 'Formatting'                  >> beam.Map(lambda (table_name, rows): (table_name, list(rows))))        # have to force to list here (dataflowRunner produces _Unwindowedvalues)

  iso_datetime = dt.datetime.now().replace(microsecond=0).isoformat()
  rows             | 'Store in GCS'                >> WriteToText(file_path_prefix=OUTPUT + iso_datetime, file_name_suffix='')

  result = p.run()
  result.wait_until_finish()

if __name__ == '__main__':
  logging.getLogger().setLevel(logging.INFO)
  run()