Skip to content

Latest commit

 

History

History
127 lines (107 loc) · 4.83 KB

File metadata and controls

127 lines (107 loc) · 4.83 KB

ETL Pipeline on GCP. PubSub publisher which simulates IoT sensors streaming data. PubSub subscriber as an ingress. DataFlow as data transform. BigQuery as a sink.

Screenshot (created with https://online.visual-paradigm.com)

as per https://github.com/GoogleCloudPlatform/training-data-analyst/tree/master/courses/streaming
and per https://www.udemy.com/gcp-data-engineer-and-cloud-architect/learn/v4/t/lecture/7598768?start=0
and per https://www.udemy.com/gcp-data-engineer-and-cloud-architect/learn/v4/t/lecture/7598772?start=0
and per https://www.udemy.com/gcp-data-engineer-and-cloud-architect/learn/v4/t/lecture/7598774?start=0 \

publish

./init.sh
./run.sh
chmod u+x ...
holds sensors simulator publisher
data is from sensors along San Diego highway
sensors publish speeds of cars in a particular lane
speedFactor 60 sends roughly 477 events every 5 seconds
nb: deeper dive at https://github.com/GoogleCloudPlatform/python-docs-samples/blob/master/pubsub/cloud-client/subscriber.py
https://github.com/GoogleCloudPlatform/python-docs-samples/blob/master/pubsub/cloud-client/publisher.py

sensors' data format:
TIMESTAMP,LATITUDE,LONGITUDE,FREEWAY_ID,FREEWAY_DIR,LANE,SPEED
2008-11-01 00:00:00,32.749679,-117.155519,163,S,1,71.2

sensors.py

publish() publishes messages
see commit a8b30485f6ec4f834efd30a1b69155daf530d74d for obsolete batch msging
note TOPIC const
simulate() simulates sensors data
-compute_sleep_secs() determines how long to wait based on speedFactor
-when time passes publish() is called

init.sh

fetches sensors data

subscribe.py

an illustration of how the stream could be consumed
consumes msgs from pull sub
effectively acts as push sub
run: python subscribe.py --project=$DEVSHELL_PROJECT_ID --topic=sensors --name=sensorsSub \

pub Screenshot sub Screenshot

process

./init.sh iot.sensors
./run.sh $DEVSHELL_PROJECT_ID iot.sensors sensors
chmod u+x ...
consumes PubSub topic stream in DataFlow calculates average speed on each highway
sinks to BigQuery

init.sh

creates BigQuery dataset
creates BigQuery table
pips

run.sh

runs Apache Beam pipeline on Cloud DataFlow backend

pipeline.py

Apache Beam pipeline built with Python SDK
streaming connection to PubSub via ReadFromPubSub
performs the following transforms:
SpeedOnFreewayFn extracts freeway id and speed
WindowInto takes 5-minutes' windows
note: needs to be performed BEFORE grouping operation like GroupByKey otherwise will be in pending indefinetely
GroupByKey groups by freeway
resolve_average_speed gets the average speed on the freeway
FormatBQRowFn adds window start&end and formats into serializable
WriteToBigQuery sinks to BigQuery

dag Screenshot Screenshot bq Screenshot

NB

streaming for python SDK is in beta

monitor progress in gcp py sdk release notes apache beam release notes
(first 2 links below)
like the is that dataflow UI is sluggish, kinda useless to monitor
user the stackdriver logs and metrics and output itself
also some useful stuff on the streaming can be found below
https://cloud.google.com/dataflow/release-notes/release-notes-python
https://beam.apache.org/blog/2018/06/26/beam-2.5.0.html
https://github.com/apache/beam/blob/master/sdks/python/apache_beam/examples/windowed_wordcount.py
https://github.com/apache/beam/blob/master/sdks/python/apache_beam/examples/streaming_wordcount.py
https://beam.apache.org/documentation/sdks/pydoc/2.6.0/apache_beam.io.gcp.bigquery.html
https://beam.apache.org/documentation/sdks/python-streaming/
https://beam.apache.org/get-started/mobile-gaming-example/#leaderboard-streaming-processing-with-real-time-game-data
https://beam.apache.org/documentation/programming-guide/index.html#core-beam-transforms
https://cloud.google.com/dataflow/pipelines/specifying-exec-params#streaming-execution
https://cloud.google.com/blog/products/data-analytics/dataflow-stream-processing-now-supports-python
https://cloud.google.com/blog/products/data-analytics/review-of-input-streaming-connectors-for-apache-beam-and-apache-spark
https://github.com/apache/beam/blob/master/sdks/python/apache_beam/examples/cookbook/bigquery_tornadoes.py \

BigQuery streaming

you might be trapped with this table has records in the streaming buffer that may not be visible in the preview
thats about bq not updating UI
query will yield though

ParDos

lambdas removed for logging sake
uncomment to revert

verbosity

uncomment to make verbose
then select stackdriver logging

stackdriver Screenshot Screenshot Screenshot

explore

contains jupyter notebook for gcp DataLab
few basics data explorations
using BigQuery and Charts API

Screenshot