by Jasmine Omeke, Obi-Ike Nwoke, Olek Gorajek
This submit is for all information practitioners, who’re serious about studying about bootstrapping, standardization and automation of batch information pipelines at Netflix.
You might keep in mind Dataflow from the submit we wrote final 12 months titled Information pipeline asset administration with Dataflow. That article was a deep dive into one of many extra technical points of Dataflow and didn’t correctly introduce this software within the first place. This time we’ll attempt to give justice to the intro after which we are going to concentrate on one of many very first options Dataflow got here with. That function is known as pattern workflows, however earlier than we begin in let’s have a fast have a look at Dataflow normally.
Dataflow
Dataflow is a command line utility constructed to enhance expertise and to streamline the info pipeline growth at Netflix. Take a look at this excessive degree Dataflow assist command output beneath:
$ dataflow --help
Utilization: dataflow [OPTIONS] COMMAND [ARGS]...Choices:
--docker-image TEXT Url of the docker picture to run in.
--run-in-docker Run dataflow in a docker container.
-v, --verbose Permits verbose mode.
--version Present the model and exit.
--help Present this message and exit.
Instructions:
migration Handle schema migration.
mock Generate or validate mock datasets.
undertaking Handle a Dataflow undertaking.
pattern Generate absolutely useful pattern workflows.
As you possibly can see Dataflow CLI is split into 4 predominant topic areas (or instructions). Probably the most generally used one is dataflow undertaking, which helps people in managing their information pipeline repositories by way of creation, testing, deployment and few different actions.
The dataflow migration command is a particular function, developed single handedly by Stephen Huenneke, to totally automate the communication and monitoring of an information warehouse desk adjustments. Due to the Netflix inner lineage system (constructed by Girish Lingappa) Dataflow migration can then aid you establish downstream utilization of the desk in query. And at last it could possibly aid you craft a message to all of the homeowners of those dependencies. After your migration has began Dataflow may also maintain observe of its progress and aid you talk with the downstream customers.
Dataflow mock command is one other standalone function. It helps you to create YAML formatted mock information information primarily based on chosen tables, columns and some rows of knowledge from the Netflix information warehouse. Its predominant objective is to allow straightforward unit testing of your information pipelines, however it could possibly technically be utilized in some other conditions as a readable information format for small information units.
All of the above instructions are very prone to be described in separate future weblog posts, however proper now let’s concentrate on the dataflow pattern command.
Dataflow pattern workflows is a set of templates anybody can use to bootstrap their information pipeline undertaking. And by “pattern” we imply “an instance”, like meals samples in your native grocery retailer. One of many predominant causes this function exists is rather like with meals samples, to present you “a style” of the manufacturing high quality ETL code that you would encounter contained in the Netflix information ecosystem.
All of the code you get with the Dataflow pattern workflows is absolutely useful, adjusted to your atmosphere and remoted from different pattern workflows that others generated. This pipeline is secure to run the second it reveals up in your listing. It should, not solely, construct a pleasant instance combination desk and fill it up with actual information, however it can additionally current you with an entire set of really helpful elements:
- clear DDL code,
- correct desk metadata settings,
- transformation job (in a language of selection) wrapped in an elective WAP (Write, Audit, Publish) sample,
- pattern set of knowledge audits for the generated information,
- and a completely useful unit take a look at in your transformation logic.
And final, however not least, these pattern workflows are being examined constantly as a part of the Dataflow code change protocol, so you possibly can make sure that what you get is working. That is one technique to construct belief with our inner consumer base.
Subsequent, let’s take a look on the precise enterprise logic of those pattern workflows.
Enterprise Logic
There are a number of variants of the pattern workflow you will get from Dataflow, however all of them share the identical enterprise logic. This was a aware determination with a purpose to clearly illustrate the distinction between varied languages wherein your ETL may very well be written in. Clearly not all instruments are made with the identical use case in thoughts, so we’re planning so as to add extra code samples for different (than classical batch ETL) information processing functions, e.g. Machine Studying mannequin constructing and scoring.
The instance enterprise logic we use in our template computes the highest hundred films/reveals in each nation the place Netflix operates every day. This isn’t an precise manufacturing pipeline operating at Netflix, as a result of it’s a extremely simplified code nevertheless it serves nicely the aim of illustrating a batch ETL job with varied transformation levels. Let’s overview the transformation steps beneath.
Step 1: every day, incrementally, sum up all viewing time of all films and reveals in each nation
WITH STEP_1 AS (
SELECT
title_id
, country_code
, SUM(view_hours) AS view_hours
FROM some_db.source_table
WHERE playback_date = CURRENT_DATE
GROUP BY
title_id
, country_code
)
Step 2: rank all titles from most watched to least in each county
WITH STEP_2 AS (
SELECT
title_id
, country_code
, view_hours
, RANK() OVER (
PARTITION BY country_code
ORDER BY view_hours DESC
) AS title_rank
FROM STEP_1
)
Step 3: filter all titles to the highest 100
WITH STEP_3 AS (
SELECT
title_id
, country_code
, view_hours
, title_rank
FROM STEP_2
WHERE title_rank <= 100
)
Now, utilizing the above easy 3-step transformation we are going to produce information that may be written to the next Iceberg desk:
CREATE TABLE IF NOT EXISTS ${TARGET_DB}.dataflow_sample_results (
title_id INT COMMENT "Title ID of the film or present."
, country_code STRING COMMENT "Nation code of the playback session."
, title_rank INT COMMENT "Rank of a given title in a given nation."
, view_hours DOUBLE COMMENT "Complete viewing hours of a given title in a given nation."
)
COMMENT
"Instance dataset delivered to you by Dataflow. For extra data on this
and different examples please go to the Dataflow documentation web page."
PARTITIONED BY (
date DATE COMMENT "Playback date."
)
STORED AS ICEBERG;
As you possibly can infer from the above desk construction we’re going to load about 19,000 rows into this desk every day. And they’re going to look one thing like this:
sql> SELECT * FROM foo.dataflow_sample_results
WHERE date = 20220101 and country_code = 'US'
ORDER BY title_rank LIMIT 5;title_id | country_code | title_rank | view_hours | date
----------+--------------+------------+------------+----------
11111111 | US | 1 | 123 | 20220101
44444444 | US | 2 | 111 | 20220101
33333333 | US | 3 | 98 | 20220101
55555555 | US | 4 | 55 | 20220101
22222222 | US | 5 | 11 | 20220101
(5 rows)
With the enterprise logic out of the best way, we are able to now begin speaking in regards to the elements, or the boiler-plate, of our pattern workflows.
Elements
Let’s take a look at the most typical workflow elements that we use at Netflix. These elements might not match into each ETL use case, however are used usually sufficient to be included in each template (or pattern workflow). The workflow writer, in spite of everything, has the ultimate phrase on whether or not they wish to use all of those patterns or maintain just some. Both means they’re right here to begin with, able to go, if wanted.
Workflow Definitions
Beneath you possibly can see a typical file construction of a pattern workflow bundle written in SparkSQL.
.
├── backfill.sch.yaml
├── each day.sch.yaml
├── predominant.sch.yaml
├── ddl
│ └── dataflow_sparksql_sample.sql
└── src
├── mocks
│ ├── dataflow_pyspark_sample.yaml
│ └── some_db.source_table.yaml
├── sparksql_write.sql
└── test_sparksql_write.py
Above bolded information outline a collection of steps (a.okay.a. jobs) their cadence, dependencies, and the sequence wherein they need to be executed.
That is a technique we are able to tie elements collectively right into a cohesive workflow. In each pattern workflow bundle there are three workflow definition information that work collectively to supply versatile performance. The pattern workflow code assumes a each day execution sample, however it is rather straightforward to regulate them to run at totally different cadence. For the workflow orchestration we use Netflix homegrown Maestro scheduler.
The predominant workflow definition file holds the logic of a single run, on this case one day-worth of knowledge. This logic consists of the next elements: DDL code, desk metadata data, information transformation and some audit steps. It’s designed to run for a single date, and meant to be referred to as from the each day or backfill workflows. This predominant workflow will also be referred to as manually throughout growth with arbitrary run-time parameters to get a really feel for the workflow in motion.
The each day workflow executes the predominant one every day for the predefined variety of earlier days. That is typically essential for the aim of catching up on some late arriving information. That is the place we outline a set off schedule, notifications schemes, and replace the “excessive water mark” timestamps on our goal desk.
The backfill workflow executes the predominant for a specified vary of days. That is helpful for restating information, most frequently due to a metamorphosis logic change, however typically as a response to upstream information updates.
DDL
Typically, step one in an information pipeline is to outline the goal desk construction and column metadata through a DDL assertion. We perceive that some people select to have their output schema be an implicit results of the remodel code itself, however the express assertion of the output schema just isn’t solely helpful for including desk (and column) degree feedback, but additionally serves as one technique to validate the remodel logic.
.
├── backfill.sch.yaml
├── each day.sch.yaml
├── predominant.sch.yaml
├── ddl
│ └── dataflow_sparksql_sample.sql
└── src
├── mocks
│ ├── dataflow_pyspark_sample.yaml
│ └── some_db.source_table.yaml
├── sparksql_write.sql
└── test_sparksql_write.py
Usually, we favor to execute DDL instructions as a part of the workflow itself, as a substitute of operating outdoors of the schedule, as a result of it simplifies the event course of. See beneath instance of hooking the desk creation SQL file into the predominant workflow definition.
- job:
id: ddl
sort: Spark
spark:
script: $S3{./ddl/dataflow_sparksql_sample.sql}
parameters:
TARGET_DB: ${TARGET_DB}
Metadata
The metadata step gives context on the output desk itself in addition to the info contained inside. Attributes are set through Metacat, which is a Netflix inner metadata administration platform. Beneath is an instance of plugging that metadata step within the predominant workflow definition
- job:
id: metadata
sort: Metadata
metacat:
tables:
- ${CATALOG}/${TARGET_DB}/${TARGET_TABLE}
proprietor: ${username}
tags:
- dataflow
- pattern
lifetime: 123
column_types:
date: pk
country_code: pk
rank: pk
Transformation
The transformation step (or steps) will be executed within the developer’s language of selection. The instance beneath is utilizing SparkSQL.
.
├── backfill.sch.yaml
├── each day.sch.yaml
├── predominant.sch.yaml
├── ddl
│ └── dataflow_sparksql_sample.sql
└── src
├── mocks
│ ├── dataflow_pyspark_sample.yaml
│ └── some_db.source_table.yaml
├── sparksql_write.sql
└── test_sparksql_write.py
Optionally, this step can use the Write-Audit-Publish sample to make sure that information is appropriate earlier than it’s made out there to the remainder of the corporate. See instance beneath:
- template:
id: wap
sort: wap
tables:
- ${CATALOG}/${DATABASE}/${TABLE}
write_jobs:
- job:
id: write
sort: Spark
spark:
script: $S3{./src/sparksql_write.sql}
Audits
Audit steps will be outlined to confirm information high quality. If a “blocking” audit fails, the job will halt and the write step just isn’t dedicated, so invalid information won’t be uncovered to customers. This step is elective and configurable, see a partial instance of an audit from the predominant workflow beneath.
data_auditor:
audits:
- operate: columns_should_not_have_nulls
blocking: true
params:
desk: ${TARGET_TABLE}
columns:
- title_id
…
Excessive-Water-Mark Timestamp
A profitable write will usually be adopted by a metadata name to set the legitimate time (or high-water mark) of a dataset. This enables different processes, consuming our desk, to be notified and begin their processing. See an instance excessive water mark job from the predominant workflow definition.
- job:
id: hwm
sort: HWM
metacat:
desk: ${CATALOG}/${TARGET_DB}/${TARGET_TABLE}
hwm_datetime: ${EXECUTION_DATE}
hwm_timezone: ${EXECUTION_TIMEZONE}
Unit Checks
Unit take a look at artifacts are additionally generated as a part of the pattern workflow construction. They consist of knowledge mocks, the precise take a look at code, and a easy execution harness relying on the workflow language. See the bolded file beneath.
.
├── backfill.sch.yaml
├── each day.sch.yaml
├── predominant.sch.yaml
├── ddl
│ └── dataflow_sparksql_sample.sql
└── src
├── mocks
│ ├── dataflow_pyspark_sample.yaml
│ └── some_db.source_table.yaml
├── sparksql_write.sql
└── test_sparksql_write.py
These unit exams are meant to check one “unit” of knowledge remodel in isolation. They are often run throughout growth to shortly seize code typos and syntax points, or throughout automated testing/deployment section, to guarantee that code adjustments haven’t damaged any exams.
We wish unit exams to run shortly in order that we are able to have steady suggestions and quick iterations through the growth cycle. Operating code in opposition to a manufacturing database will be sluggish, particularly with the overhead required for distributed information processing techniques like Apache Spark. Mocks mean you can run exams regionally in opposition to a small pattern of “actual” information to validate your transformation code performance.
Languages
Over time, the extraction of knowledge from Netflix’s supply techniques has grown to embody a wider vary of end-users, similar to engineers, information scientists, analysts, entrepreneurs, and different stakeholders. Specializing in comfort, Dataflow permits for these differing personas to go about their work seamlessly. A lot of our information customers make use of SparkSQL, pyspark, and Scala. A small however rising contingency of knowledge scientists and analytics engineers use R, backed by the Sparklyr interface or different information processing instruments, like Metaflow.
With an understanding that the info panorama and the applied sciences employed by end-users usually are not homogenous, Dataflow creates a malleable path towards. It solidifies totally different recipes or repeatable templates for information extraction. Inside this part, we’ll preview just a few strategies, beginning with sparkSQL and python’s method of making information pipelines with dataflow. Then we’ll segue into the Scala and R use circumstances.
To start, after putting in Dataflow, a consumer can run the next command to grasp get began.
$ dataflow pattern workflow --help
Dataflow (0.6.16)Utilization: dataflow pattern workflow [OPTIONS] RECIPE [TARGET_PATH]
Create a pattern workflow primarily based on chosen RECIPE and land it within the
specified TARGET_PATH.
At present supported workflow RECIPEs are: spark-sql, pyspark,
scala and sparklyr.
If TARGET_PATH:
- if not specified, present listing is assumed
- factors to a listing, will probably be used because the goal location
Choices:
--source-path TEXT Supply path of the pattern workflows.
--workflow-shortname TEXT Workflow quick title.
--workflow-id TEXT Workflow ID.
--skip-info Skip the information in regards to the workflow pattern.
--help Present this message and exit.
As soon as once more, let’s assume we now have a listing referred to as stranger-data wherein the consumer creates workflow templates in all 4 languages that Dataflow affords. To higher illustrate generate the pattern workflows utilizing Dataflow, let’s have a look at the total command one would use to create one in every of these workflows, e.g:
$ cd stranger-data
$ dataflow pattern workflow spark-sql ./sparksql-workflow
By repeating the above command for every sort of transformation language we are able to arrive on the following listing construction
.
├── pyspark-workflow
│ ├── predominant.sch.yaml
│ ├── each day.sch.yaml
│ ├── backfill.sch.yaml
│ ├── ddl
│ │ └── ...
│ ├── src
│ │ └── ...
│ └── tox.ini
├── scala-workflow
│ ├── construct.gradle
│ └── ...
├── sparklyR-workflow
│ └── ...
└── sparksql-workflow
└── ...
Earlier we talked in regards to the enterprise logic of those pattern workflows and we confirmed the Spark SQL model of that instance information transformation. Now let’s focus on totally different approaches to writing the info in different languages.
PySpark
This partial pySpark code beneath may have the identical performance because the SparkSQL instance above, nevertheless it makes use of Spark dataframes Python interface.
def predominant(args, spark):source_table_df = spark.desk(f"{some_db}.{source_table})
viewing_by_title_country = (
source_table_df.choose("title_id", "country_code",
"view_hours")
.filter(col("date") == date)
.filter("title_id IS NOT NULL AND view_hours > 0")
.groupBy("title_id", "country_code")
.agg(F.sum("view_hours").alias("view_hours"))
)
window = Window.partitionBy(
"country_code"
).orderBy(col("view_hours").desc())
ranked_viewing_by_title_country = viewing_by_title_country.withColumn(
"title_rank", rank().over(window)
)
ranked_viewing_by_title_country.filter(
col("title_rank") <= 100
).withColumn(
"date", lit(int(date))
).choose(
"title_id",
"country_code",
"title_rank",
"view_hours",
"date",
).repartition(1).write.byName().insertInto(
target_table, overwrite=True
)
Scala
Scala is one other Dataflow supported recipe that provides the identical enterprise logic in a pattern workflow out of the field.
bundle com.netflix.sparkobject ExampleApp {
import spark.implicits._
def readSourceTable(sourceDb: String, dataDate: String): DataFrame =
spark
.desk(s"$someDb.source_table")
.filter($"playback_start_date" === dataDate)
def viewingByTitleCountry(sourceTableDF: DataFrame): DataFrame = {
sourceTableDF
.choose($"title_id", $"country_code", $"view_hours")
.filter($"title_id".isNotNull)
.filter($"view_hours" > 0)
.groupBy($"title_id", $"country_code")
.agg(F.sum($"view_hours").as("view_hours"))
}
def addTitleRank(viewingDF: DataFrame): DataFrame = {
viewingDF.withColumn(
"title_rank", F.rank().over(
Window.partitionBy($"country_code").orderBy($"view_hours".desc)
)
)
}
def writeViewing(viewingDF: DataFrame, targetTable: String, dataDate: String): Unit = {
viewingDF
.choose($"title_id", $"country_code", $"title_rank", $"view_hours")
.filter($"title_rank" <= 100)
.repartition(1)
.withColumn("date", F.lit(dataDate.toInt))
.writeTo(targetTable)
.overwritePartitions()
}
def predominant():
sourceTableDF = readSourceTable("some_db", "source_table", 20200101)
viewingDf = viewingByTitleCountry(sourceTableDF)
titleRankedDf = addTitleRank(viewingDF)
writeViewing(titleRankedDf)
R / sparklyR
As Netflix has a rising cohort of R customers, R is the most recent recipe out there in Dataflow.
suppressPackageStartupMessages({
library(sparklyr)
library(dplyr)
})...
predominant <- operate(args, spark) {
title_df <- tbl(spark, g("{some_db}.{source_table}"))
title_activity_by_country <- title_df |>
filter(title_date == date) |>
filter(!is.null(title_id) & event_count > 0) |>
choose(title_id, country_code, event_type) |>
group_by(title_id, country_code) |>
summarize(event_count = sum(event_type, na.rm = TRUE))
ranked_title_activity_by_country <- title_activity_by_country |>
group_by(country_code) |>
mutate(title_rank = rank(desc(event_count)))
top_25_title_by_country <- ranked_title_activity_by_country |>
ungroup() |>
filter(title_rank <= 25) |>
mutate(date = as.integer(date)) |>
choose(
title_id,
country_code,
title_rank,
event_count,
date
)
top_25_title_by_country |>
sdf_repartition(partitions = 1) |>
spark_insert_table(target_table, mode = "overwrite")
}
predominant(args = args, spark = spark)
}
