쉐도잉 연습: Data Ingestion Best Practices! How to Ingest Data Efficiently at Enterprise Scale! - 영상으로 영어 말하기 배우기

Hey y'all, Data Guy here and today I'm back with yet another viewer request video,

this time talking about best practices for data ingestion.

So really focusing on the first part of your data pipelines, really the first part of almost any data engineering workflow,

right, is bringing data into the platform that you want that data to be brought into.

And so that's what we're going to explore today, is really how you can design effective systems for data ingestion regardless of what tools you're using.

So this isn't really a video exploring, hey, what tools you use for data ingestion, but rather what high level design practices, which are really global to any kind of data ingestion use case,

should you apply

when you're designing your data ingestion workflows to make sure you're designing them as efficiently as possible

and you're designing them for enterprise scale because that's what everyone wants, right?

So if you like these videos, please like, subscribe, consider joining my Patreon, It helps me out a ton, but without further ado, let's get into it.

Now, when you're initially designing your data pipelines and data ingestion, you want to design for scale and elasticity above all else.

Especially in cloud architectures, it is a necessity because data volumes can spark unpredictably due to external events, user behavior, system interactions.

There's so many different ways that your data volumes can spike unexpectedly.

And so your best practice here is going to be using things that are serverless or auto-scaling.

So, you know, servers like AWS Lambda functions or Google Cloud functions can ingest bursts of data without provisioning infrastructure,

but really any kind of serverless or auto-scaling component that you might want to use to, or

that you're going to need to use to make sure you

can handle those bursts of data without having a bunch of extra compute sitting idly by.

And in tandem with this, managed integration services like AWS Kinesis

or Google PubSub also help provide built-in scalability where you only pay for what you use

and so they're designed for high throughput ingestion in that way

because they'll automatically scale to match the amount of data that you actually need to adjust.

And another thing to you know think about designing for even

at the earlier stages of your data ingestion pipelines are using time-based

or key-based partition strategies to in your object storage systems to prevent any performance bottlenecks downstream,

you know, by making your data partitioned effectively earlier on, then the downstream queries don't need to query and go over all of that data.

They only need to go toward through their relevant partition.

Now, the next thing you want to think about is ensuring data quality at your ingestion point.

It's a lot cheaper to catch errors early and trim bad data than processing it and trying to fix it downstream.

Bad data is going to create bad outputs.

And so by not putting garbage in, you're not gonna get garbage out.

And so to accomplish this, you want to do things like input data quality checks to validate schemas at ingestion, you know, using schema enforcement tools.

You know, there's glue schema registry, there's confluent schema registry that you can use to compare the ingested schema versus an existing schema.

And there's also open source tools as well to make sure that the data is of the expected shape.

And then also implement data contracts, which are basically strict expectations programmatically defined for incoming data from producers using tools like open API,

protocol buffers, or even just having JSON schemas that say this data has to match this particular format.

And then also, your biggest kind of guard against bad data is leveraging integrating tools like Soda,

like Great Expectations, to automatically verify and apply more complex freshness

or no value checks or to check distribution of data or type consistency.

Those tools will actually help you dig into the data

and understand it rather than just doing high level schema or shape validation.

So it's a combination of both of those tools using conjunction

to help raise any issues as early as possible in the data pipeline to stop that bad data getting processed any further.

Now, the next thing you're going to want to think about, especially in today's day and age, is being able to support both batch and streaming workloads.

It's highly unlikely in modern stacks that you're only going to be doing batch processing, only going to be doing stream processing,

and you're likely going to be doing some kind of hybrid pipeline where, you know, you might be processing historical data in bulk and then streaming any real-time updates after that fact, right?

And the best practices here for making sure you can support

both of these use cases are to use cloud-native hybrid ingestion frameworks, you know, Databricks structured streaming and,

you know, Spark structured streaming essentially is a great way to, hey, if I want to use Spark for both streaming and batch, you can do that.

You also have Flink, Apache Flink, which will run both batch and streaming workloads or Apache Beam as well.

And within here as well, you know, other than, you know, the kind of tools and platforms you want to use, you also are going to want to design your ingestion logic for item potency

and making sure that your ingestion logic can safely replay

or reprocess data and get the same result to avoid duplication and loss,

because that's really critical with streaming workloads and avoiding data from being processed twice and again gumming up your backend databases.

Another good practice here is to separate raw and processed stones.

You have a raw data ingestion zone in S3

or GCS bucket before anything actually gets processed to support reprocessing

and recovery if you ever need to go back and reprocess any raw data as well.

Those are all great ways to make sure you're able to accomplish batch and streaming use cases equally effectively.

Now the next topic I want to discuss is everyone's favorite and that is secured.

Data ingestion is most often the way you're going to be ingesting and opening your systems up to sensitive data,

and you're going to need to implement proper protections and observability on that to make sure it stays secure.

And the first thing, you're going to want to make sure your data is always encrypted in transit and at rest.

So use things like TLS for all network transmission and cloud-native encryption protocols for stored data.

and then also implement access controls using your IAM policies to restrict access by source, type, and role and implement least privileged based access

so everyone is only scoped to access the data that they are explicitly need for their job, right?

And then enable, you know, observability across your entire pipeline, you know, make sure your ingestion pipelines have structured logging

and metrics and tracing using tools like Prometheus and OpenTelemetry and Datadog

So you can also set up further gates against kind of bad data or, you know, any kind of unexpected or anomalous behavior, which might indicate, you know,

some type of, you know, leaking of data or, you know, some kind of degradation in your security.

Now, the next thing I'll talk about is designing for fault tolerance and resiliency.

A big piece of designing good, effective data investment pipelines is making sure they can gracefully handle failures to maintain SLAs and data integrity.

And here, you know, number one, you're going to want to think about ensuring a high availability, deploy ingestion services across multiple zones or regions when possible.

You know, things like multi-zone Kafka clusters or regional failover in Event Hub.

So if there's a failure or degradation service in one region, your jobs will start to get processed in another active region.

Another thing you might want to use is something like a dead letter queue.

And that will route any unprocessable records to separate storage instead of just deleting them

or throwing them out for later inspection remediation so

if you get bad data you can actually look at

that bad data down the line understand hey why did we get this

and fix the process or the issue that caused

that bad data in the first place

and then also retry logic making sure you're have retry logical exponential back off

so you don't flood systems during outages now

if you don't have exponential back off you know

and you just have retry logic then all of your pipelines

that are failing are going to start trying again and pinging

and just going to overload any back-end service that they're connecting to.

So you want to make sure that you have exponential backoffs

that gets spread out over a longer time period and you don't have those service overloads occur.

Now, the next thing you want to think about is metadata capture and lineage tracking.

Metadata and lineage are vital for any kind of debugging or governance or auditing use cases.

And it's really hard to get that down the line if you aren't capturing it during ingestion.

So capture attributes like source system, like the time stand, like the transformation status of every step of your data pipeline

so that you understand how data is processed and evolved.

And then also implement lineage tracking tools like OpenLineage with Marques

or Dataplex to record and track and visualize how data flows go from start to finish.

You have a visual representation.

You can see how data is changing and evolving

100

and going through your various systems so you can identify bottlenecks or areas of issue.

101

And then also adopt a data catalog.

102

You know, you're going to want to make sure you have a central

103

place to store all of this metadata to make sure it's easily queryable and understandable and has all the context necessary.

104

And here you have, you know, something like AWS Glue catalog or Azure Purview or Google Data Catalog for discoverability and compliance and storing all this metadata.

105

Now, finally, the last piece of advice I have for you is making sure you're choosing the right tool for the job.

106

different ingestion scenarios are gonna require different tools and

107

so it's you're gonna want to make sure you're matching the right technology to your needs

108

and not trying to cram everything in one

109

that is the biggest cause of pain I see out there

110

is hey my company said we're all going all in on Databricks

111

and Spark so everything is gonna happen through Spark

112

and there's a lot of things

113

that aren't really well designed for running with Spark right and

114

so making sure you're using the right tool for each job

115

and having a whole different set of different tools for each scenario you might need to encounter is crucial

116

And it doesn't and won't force you to just cram use cases into tools

117

that aren't designed for that and have them run in efficiently.

118

So some general guidelines, you can see some examples in here, is for any kind of real-time streaming tool,

119

use a dedicated real-time streaming tool like Kafka or Pulsar or Kinesis or Flink or Azure Event Hub or Cloud PubSub.

120

For tools like for change data capture, that's where tools like Debezium is probably the gold standard.

121

Then you also have 5Tran or Airbyte that can also stream changes from operational databases as well.

122

And then for any kind of batch data ingestion, orchestrators like Airflow or Mage

123

or Daxter are really great tools for managing all the different stages of your data pipeline

124

and managing all the different tools you might need to ingest from instead of learning mechanisms

125

and give you a single kind of framework to manipulate data and move data around your environments.

126

And then finally, for IoT or event heavy pipelines, then you're actually going to want to look at time series ingestion platforms like InfluxDB or TimescaleDB,

127

which are going to be really well custom designed for those very specific set of needs you use for those use cases.

128

So that is really all different things I wanted to talk about here.

129

Data ingestion in cloud architectures is no longer, hey, you can just have a one size fits all approach.

130

you need to have modularity, elasticity, flexibility to design effective data pipelines in the day, state, and age.

131

So I hope this video has helped you figure out how to do that.

132

I hope you have a great rest of your day.

133

Data Guy out.

쉐도잉이란? 영어 실력을 빠르게 키우는 과학적 방법

쉐도잉(Shadowing)은 원래 전문 통역사 훈련을 위해 개발된 언어 학습 기법으로, 다언어 학자인 Dr. Alexander Arguelles에 의해 대중화된 방법입니다. 핵심 원리는 간단하지만 매우 강력합니다: 원어민의 영어를 들으면서 1~2초의 짧은 지연으로 즉시 소리 내어 따라 말하는 것——마치 '그림자(shadow)'처럼 화자를 따라가는 것입니다. 문법 공부나 수동적인 청취와 달리, 쉐도잉은 뇌와 입 근육이 동시에 실시간으로 영어를 처리하고 재현하도록 훈련합니다. 연구에 따르면 이 방법은 발음 정확도, 억양, 리듬, 연음, 청취력, 말하기 유창성을 크게 향상시킵니다. IELTS 스피킹 준비와 자연스러운 영어 소통을 원하는 분들에게 특히 효과적입니다.

쉐도잉 연습: Data Ingestion Best Practices! How to Ingest Data Efficiently at Enterprise Scale! - 영상으로 영어 말하기 배우기

이 비디오로 말하기 연습하는 이유는?

문맥 속의 문법 및 표현

일반적인 발음 함정

쉐도잉이란? 영어 실력을 빠르게 키우는 과학적 방법