Menu

ETL, ELT, and Data Pipelines

Lesson 2: ETL, ELT, and Data Pipelines

What is a data pipeline?

A data pipeline moves data from sources, through processing steps, to a destination where it can be analysed.

Trigger type

Example

Scheduled (batch)Every night at 2 AM
Event-drivenWhen a new file lands in storage
Streaming (real-time)Continuously, as events happen

ETL and ELT

Aspect

ETL (Extract → Transform → Load)

ELT (Extract → Load → Transform)

Where transformation happensBefore loading, in a separate processing layerInside the destination warehouse
Raw data kept?Often notYes, raw data is loaded first
Typical era / settingTraditional on-premise warehousesModern cloud warehouses
FlexibilityLess: must re-extract to change logicMore: re-transform from stored raw data
Common toolsInformatica, SSIS, custom scriptsFivetran / Airbyte (load) + dbt (transform)

The right approach depends on architecture, governance, scale, and cost.

Modern data stack (common pattern):

Layer

Purpose

Example tools

IngestionExtract and loadFivetran, Airbyte
StorageWarehouse / lakehouseSnowflake, BigQuery, Databricks
TransformationSQL-based modellingdbt
OrchestrationSchedule and coordinateApache Airflow, Dagster, Prefect
BIReportingPower BI, Tableau, Looker

Example workflow

Stage

Input

Action

Output

1. ExtractOrder databasePull yesterday's ordersRaw order records
2. ValidateRaw recordsCheck fields and keysValidated records + error log
3. StandardiseValidated recordsConvert timestamps to IST, unify formatsClean records
4. CalculateClean recordsCompute agreed metricsMetrics table
5. LoadMetrics tableWrite to warehouseUpdated reporting table
6. RefreshReporting tableRefresh dashboardUpdated dashboard

Each stage should have clear inputs and outputs.

Reliability checks

Check

Example rule

Row countWithin ±20% of the 7-day average
FreshnessLatest record less than 24 hours old
UniquenessNo duplicate order IDs
CompletenessRequired fields not null
DistributionAverage order value within the expected range
Referential integrityAll product IDs exist in the product table

Alert the owner when a critical check fails, instead of silently publishing unreliable numbers. Tools such as dbt tests and Great Expectations automate these checks.

Common mistakes

  • "The job succeeded" ≠ "the data is correct." A pipeline can run perfectly and still produce wrong numbers.
  • Monitor quality and meaning of the output, not just job completion.
  • No ownership: nobody knows who fixes a failure.
  • Hard-coded dates or credentials in pipeline code.

Data Analytics Tutorial



Learn via Video Course

Data Analysis and Processing with Python using Pandas - Data Science Course Logo

Data Analysis and Processing with Python using Pandas - Data Science Course

5 hrs



Get in Touch For Details