Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Productionizing Big Data - stories from the tre...
Search
Roksolana
September 14, 2023
Technology
0
69
Productionizing Big Data - stories from the trenches
Presented at ScalaDays 2023 (Madrid, Spain)
Roksolana
September 14, 2023
Tweet
Share
More Decks by Roksolana
See All by Roksolana
Pain of engineering management
roksolanad
1
76
Alice and the return to the world of pods and higher-order functions
roksolanad
0
190
Modern data pipelines in AdTech - life in the trenches
roksolanad
1
290
Alice and travelling back in time
roksolanad
0
170
Big Data at AdTech
roksolanad
0
340
Alice and the Mad Hatter: Predict or not to predict
roksolanad
0
190
Alice in the world of machine learning
roksolanad
0
120
Alice and the lost pod: practical guide to Kubernetes in Scala
roksolanad
1
340
Scala meets Kubernetes
roksolanad
0
500
Other Decks in Technology
See All in Technology
PythonとLLMで挑む、 4コマ漫画の構造化データ化
esuji5
1
120
Optuna DashboardにおけるPLaMo2連携機能の紹介 / PFN LLM セミナー
pfn
PRO
1
820
LLMアプリケーション開発におけるセキュリティリスクと対策 / LLM Application Security
flatt_security
7
1.7k
ユニットテストに対する考え方の変遷 / Everyone should watch his live coding
mdstoy
0
110
実装で解き明かす並行処理の歴史
zozotech
PRO
1
260
OCI Network Firewall 概要
oracle4engineer
PRO
1
7.7k
多野優介
tanoyusuke
1
170
SoccerNet GSRの紹介と技術応用:選手視点映像を提供するサッカー作戦盤ツール
mixi_engineers
PRO
1
160
Go Conference 2025: GoのinterfaceとGenericsの内部構造と進化 / Go type system internals
ryokotmng
3
590
extension 現場で使えるXcodeショートカット一覧
ktombow
0
190
Trust as Infrastructure
bcantrill
0
280
関係性が駆動するアジャイル──GPTに人格を与えたら、対話を通してふりかえりを習慣化できた話
mhlyc
0
130
Featured
See All Featured
jQuery: Nuts, Bolts and Bling
dougneiner
64
7.9k
The Power of CSS Pseudo Elements
geoffreycrofte
79
6k
Helping Users Find Their Own Way: Creating Modern Search Experiences
danielanewman
30
2.9k
Raft: Consensus for Rubyists
vanstee
139
7.1k
It's Worth the Effort
3n
187
28k
Six Lessons from altMBA
skipperchong
28
4k
Large-scale JavaScript Application Architecture
addyosmani
514
110k
Context Engineering - Making Every Token Count
addyosmani
4
160
Visualization
eitanlees
148
16k
Building a Scalable Design System with Sketch
lauravandoore
462
33k
Practical Tips for Bootstrapping Information Extraction Pipelines
honnibal
PRO
23
1.5k
Faster Mobile Websites
deanohume
310
31k
Transcript
Productionizing big data - stories from the trenches
Roksolana Diachuk •Engineering manager at Captify •Women Who Code Kyiv
Data Engineering Lead •Speaker
AdTech methodologies deliver the right content at the right time
to the right consumer AdTech
None
You have your pipelines in production What’s next?
Types of issues • Low performance • Human errors •
Data source errors
Story #1. Unlucky query
Problem Drop 13 months of user profiles
Reporting
Problem 13 months hour=22042001
Loading mechanism loader.ImpalaLoaderConfig.periodToLoad: “P5D” loader.ImpalaLoaderConfig.periodToLoad: “P13M” val minTime = currentDay.minus(config.feedPeriod)
listFiles.filter(file => file.eventDateTime isAfter minTime)
Solution loader.ImpalaLoaderConfig.periodToLoad: “P5D” loader.ImpalaLoaderConfig.periodToLoad: “P1M” loader.ImpalaLoaderConfig.periodToLoad: “P13M” …
Story #2. Missing data
Data ingestion Data from Partner X Data costs attribution Extractor
Problem XX Advertiser ID, Language, XX Device Type, …, XX
Media Cost (USD) X Advertiser ID, Language, X Device Type, …, X Media Cost (USD)
Solution • Rename old columns • Reload data for the
week
Solution val colRegex: Regex = “””X (.+)“””.r val oldNewColumnsMapping =
df.schema.collect { case oldColdName@colRegex(pattern) => (oldColName.name, (“XX “ + pattern)) } oldNewColumnsMapping.foldLeft(df) { case (data, (oldName, newName)) => data.withColumnRenamed(oldName, newName) }
XX Advertiser ID, Language, XX Device Type, …, XX Media
Cost (USD) Solution
Story #3. Divide and conquer
Problem processing_time part-*.parquet filtering aggregations created part-*.parquet
• Slow processing • Large parquet files • Failing job
that consumes lots of resources Problem
• Write new partitioned state • Run downstream jobs with
smaller states • Generate seed partition column - xxhash64(fullUrl, domain) Solution
processing_time part-*.parquet created bucket=0 part-*.parquet part-*.parquet … bucket=9 part-*.parquet part-*.parquet
processing_time part-*.parquet Solution
Story #4. Catch the evolution train
Data organisation evolution
Problem • Missing columns from the source • Impala to
Databricks migration speed • Dependency with another team • Unhappy users
Log-level data Mapper Ingestor Transformer Data costs calculator Data costs
attribution
Data costs attribution Data costs attribution Data extractor Impala loader
Data costs attribution Data extractor Impala loader Data costs attribution
Solution XX Advertiser ID, Language, XX Device Type, …, XX
Partner Currency, XX CPM Fee (USD) XX Advertiser ID, Language, XX Device Type, …, XX Media Cost (USD) 26 columns 82 columns
Solution Data extractor New ingestion job
//final step is writing the data df.write .partitionBy(“event_date”, “event_hour”) .mode(SaveMode.Overwrite)
.parquet(dstPath) Solution
Why this solution doesn’t work data_feed clicks.csv.gz views.csv.gz activity.csv.gz event_date
clicks1.parquet clicks2.parquet
Impressions Clicks Conversions Attribution data source
Solution impressions clicks conversions clicks.csv.gz views.csv.gz activity.csv.gz
Story #5. Cleanup time
Corrupted data Data from Partner X Ingestor
Corrupted data Data from Partner X Ingestor IllegalArgumentException: Can't convert
value to BinaryType data type
Solution • Adjust pipeline • Reload data for 3 days
on S3 • Relaunch Databricks autoloader
Current solution impressions videoevents conversions impressions conversions Clicks clicks videoevents
Current solution impressions conversions clicks videoevents
Better solution impressions videoevents conversions impressions conversions clicks clicks videoevents
Conclusions
2. Observability is the key 4. Plan major changes carefully
1. Set up clear expectations with stakeholders Prevention mechanisms 3. Distribute data transformation load
2. Errors can be prevented 4. Data evolution is hard
1. Data setup is always changing Conclusions 3. There are multiple approaches with different tools
None
dead_ fl owers22 roksolana-d roksolanadiachuk roksolanad My contact info