Executive Summary
Decide where transformation will run and who will maintain the connectors before comparing products. Those two answers remove most of the shortlist.
ETL and ELT data integration tools move data from applications, databases and files into a warehouse or lake, and reshape it for analysis. Microsoft's Azure Architecture Center defines extract, transform, load (ETL) as a data integration process that consolidates data from diverse sources into a unified data store, in which data is modified according to business rules using a specialized engine during the transformation phase. The Azure Architecture Center states that extract, load, transform (ELT) differs from ETL solely in where the transformation takes place: in the ELT pipeline, the transformation occurs in the target data store. The choice between them decides whose compute runs the transformation and who owns the code that defines it.
AWS states that ETL has been around since the 1970s and that traditional data warehouses required custom ETL processes for each data source. Cloud warehouses changed the default. Loading raw data first lets analysts rebuild a model later without extracting the data again. The Azure Architecture Center states that ELT only works well when the target system is powerful enough to transform the data efficiently. AWS states that it is sometimes more beneficial to use ETL to integrate with legacy databases or third-party data sources with predetermined data formats.
Most purchases now cover four motions: loading data in batches, copying database changes as they happen (change data capture, or CDC), transforming data inside the warehouse, and sending modeled data back to business applications. The Azure Architecture Center describes reverse ETL as the process of moving transformed, modeled data from analytical systems into operational tools and applications. Few tools are equally strong at all four, so test each motion you need on its own.
ETL or ELT: Where Should the Transformation Run?
The Azure Architecture Center advises choosing ETL when you need to offload heavy transformations away from a constrained target system, when complex business rules require specialized transformation engines, or when regulatory or compliance requirements mandate curated staging audits before loading.
The Azure Architecture Center lists three conditions under choosing ELT: your target system is a modern data warehouse or lakehouse with elastic compute scaling; you need to preserve raw data for exploratory analysis or future schema evolution; and transformation logic benefits from the target system's native capabilities.
Most organizations run both. A finance feed with fixed rules may be cleaned before it lands, while product and marketing data lands raw and is modeled in the warehouse. The tool decision follows from that split. A tool that only extracts and loads needs a transformation layer beside it, and an engine that transforms in flight needs capacity sized for the heaviest job.
Change data capture often decides the purchase. Copying full tables on a schedule is simple, but it reloads unchanged rows and can miss deletes between runs. Log-based CDC reads the database’s transaction log instead. Debezium describes itself as an open source distributed platform for change data capture. Debezium states that log-based CDC as implemented by Debezium ensures that all data changes are captured and produces change events with a very low delay while avoiding increased CPU usage required for frequent polling. Ask each vendor which databases it reads by log and which by query, and what load each method puts on the source.
AWS describes zero-ETL as a set of integrations that minimizes the need to build ETL data pipelines. Native replication between a cloud provider’s own databases and its warehouse can remove a pipeline for those sources. It does not reach SaaS applications or databases outside that provider, so it lowers the number of connectors you buy rather than replacing them.
The Azure Architecture Center states that, unlike ETL or ELT, which operate on datasets in scheduled batches, real-time streaming processes data as it arrives. Event streaming is a separate purchase with its own guide. This guide covers batch and change-data-capture pipelines into analytical stores.
Should you build or buy ETL & ELT Data Integration Tools?
Hand-written extraction scripts are cheap to start and expensive to keep. Each source changes its API, adds fields and enforces rate limits on its own schedule, and every change can break a pipeline that someone must fix. The buying decision is which layers to own.
A pipeline has five layers: extraction and loading, change data capture, transformation, orchestration, and activation back to applications. Some vendors sell one layer, some bundle several, and the cloud providers sell each as a separate service billed on one account.
Open-source software sits between building and buying. You avoid a license fee and can change the code, and you take on hosting, upgrades and on-call duty. Read the license as well as the price list, because some source-available licenses restrict how the software may be offered to others.
| Layer | What it answers | What to buy |
|---|---|---|
| Extract and load | How data leaves each source and lands in the warehouse or lake | A managed connector service, an open-source connector platform, or a cloud provider’s service |
| Change data capture | How inserts, updates and deletes reach the target without full reloads | Log-based replication, from an integration suite or a dedicated CDC tool |
| Transformation | Where joins, cleaning and business rules run, and who maintains them | SQL models in the warehouse, a visual pipeline builder, or an ETL engine |
| Orchestration | What runs when, in what order, and what happens on failure | A workflow scheduler, managed or self-hosted |
| Reverse ETL | How modeled data reaches CRM, marketing and support tools | Activation features in an ELT platform, or a separate activation tool |
How do you evaluate ETL & ELT Data Integration Tools?
Connector coverage comes first, and the count on a vendor’s website is the wrong measure. List your sources, then ask which ones the vendor builds and maintains, which come from a community catalog, and which you would build with a software development kit. A connector the vendor maintains is one it fixes when the source changes.
Change handling decides whether pipelines survive normal operations. Ask how each tool reacts when a column is added, renamed or retyped, when a table is dropped, and when a source deletes rows. Ask how a failed sync resumes: from a checkpoint, or from the start.
Transformation is where the business logic lives. SQL models kept in version control can be reviewed, tested and rerun like application code. Visual builders are faster for small teams and harder to review at scale. Whichever you choose, check that tests and documentation sit next to the logic they describe.
Deployment and security set where your data travels. A managed service processes data in the vendor’s cloud. A hybrid model keeps processing inside your network while the vendor runs the control plane. Check private networking, encryption with your own keys, column masking and which pricing tier includes each.
Operations and lineage decide how fast you find a broken pipeline. Look for freshness alerts per table, logs your monitoring tools can read, and lineage from source column to report.
| Capability | What it does | Buyer translation |
|---|---|---|
| Maintained connectors | Extracts from a source and keeps up with its API changes | Which of your sources the vendor fixes, and how fast |
| Log-based CDC | Reads changes from the database transaction log | Fresh data with less load on the source |
| Schema change handling | Adapts the target when the source adds or changes columns | Fewer broken pipelines after source releases |
| Transformation in the warehouse | Runs SQL models where the data already sits | Warehouse compute on your bill, logic in your repository |
| Hybrid deployment | Processes data inside your network under a managed control plane | Data stays local; you run the workers |
| Reverse ETL | Sends modeled data to business applications | One pipeline vendor or two |
Vendors in ETL & ELT Data Integration Tools
Representative vendors include Airbyte, Apache Airflow / Astronomer, AWS Glue, Fivetran + dbt Labs, Informatica / Salesforce, Matillion, Microsoft Fabric / Azure Data Factory and Qlik Talend. The list is alphabetical, is not a ranking, and is not for sale. How we choose vendors
The approaches below follow two questions: who builds and maintains the pipelines, and where does the transformation run? Several vendors sell more than one layer, so each is placed by what buyers mainly purchase it for.
Managed ELT services maintain the connectors and load raw data into your warehouse. Fivetran announced on June 1, 2026 the completion of its all-stock merger with dbt Labs, originally announced on October 13, 2025. Fivetran states that it moves data from 750+ sources and activates it back into 200+ destinations. dbt Labs states that with dbt you write simple SQL select statements, and dbt handles the heavy lifting by creating modular, maintainable data models. dbt Core v2.0 (alpha) is the open sourcing of the dbt Fusion engine runtime, released under an Apache 2.0 license. Fivetran states that an activation sync moves data from an activation source to an activation destination, and that this was previously called a sync in Census. Smaller managed services compete on pricing model and latency. Hevo Data states that Hevo combines extraction, loading, and dbt-based modeling into a single platform. Estuary states that most teams run two tools that read the same data twice, while Estuary reads it once, stores it in your own cloud, and delivers it to each destination at the speed it needs.
Open-source ELT lets you run the connector platform yourself. Airbyte states that it replicates data from 700+ sources, from legacy databases like Oracle and Db2 to SaaS apps like Salesforce, inside your VPC or on Airbyte Cloud. Airbyte states that its data replication platform is available as a self-managed, hybrid, or fully managed cloud solution. Airbyte states that Airbyte Core is free to self-host. Airbyte states that you cannot sell a cloud service that provides users with direct access to Airbyte.
Enterprise data integration suites bundle ETL, ELT, CDC, data quality and governance under one contract, and most run in the cloud and on premises. Salesforce announced that it has completed its acquisition of Informatica. Salesforce states that MuleSoft combined with Informatica’s data integration and governance capabilities will create a comprehensive, end-to-end integration offering. Informatica states that you can ingest data into cloud data warehouses and lakes with high-performance ELT/ETL, data replication or change data capture. Informatica offers PowerCenter customers a path to modernize to the cloud with its Intelligent Data Management Cloud (IDMC). Qlik states that Talend is now part of Qlik. A request for talend.com is answered by qlik.com. Qlik states that as of January 31, 2024, the open-source version of Talend Studio was retired and is no longer hosted or updated by Qlik and Talend. Qlik states that Qlik Data Integration is a client managed solution that can be installed on-premises or as a virtual machine image in any location you choose. Stitch states that Stitch is now part of Qlik.
Each major cloud provider sells integration services billed on the account you already have. AWS describes AWS Glue as a serverless data integration service to discover, prepare, and combine data for analytics, AI and agents. AWS states that AWS Glue offers flexible support for workloads like ETL, ELT, and streaming in one service. Microsoft describes Data Factory in Microsoft Fabric as the next generation of Azure Data Factory, with a simpler architecture, built-in AI, and new features. Microsoft describes Azure Data Factory as a managed cloud service built for complex hybrid extract-transform-load (ETL), extract-load-transform (ELT), and data integration projects. Microsoft states that Data Factory connects to more than 170 data sources, including multicloud environments and hybrid setups with on-premises gateways.
Visual pipeline builders target teams that design pipelines on a canvas rather than in code. Matillion states that it went to market with Matillion ETL, an on-premise solution. Matillion states that Data Productivity Cloud has been renamed to Maia Foundation, part of Maia, Matillion's AI Data Automation platform. Matillion lists 150+ connectors, orchestration, and governance across Snowflake, Databricks, BigQuery, and Redshift.
Orchestrators do not move data themselves. They schedule and sequence the jobs that do, and retry them when they fail. The Apache Airflow project describes Airflow as an open-source platform for developing, scheduling, and monitoring workflows. Apache Airflow 3.0.0 has been released. The Apache Airflow project states that in Airflow 3, a DAG will run through to completion based on the version at start, even if a new version has been uploaded while it was being run. Astronomer describes Astro as the unified orchestration platform for Apache Airflow. AWS states that Amazon Managed Workflows for Apache Airflow (MWAA) lets you orchestrate data pipelines and workflows with Apache Airflow, with serverless or provisioned deployment options. Google Cloud describes Managed Service for Apache Airflow (formerly Cloud Composer) as a fully managed workflow orchestration service built on Apache Airflow.
Application integration vendors also sell data pipelines, which suits teams that want one vendor for both. Boomi states that Boomi Data Integration provides a unified solution for Ingestion, Transformation, Orchestration, Activation, and Data Operations. Boomi states that Boomi Data Integration connects any source to any destination through no-code pipelines and real-time Change Data Capture (CDC) replication. SnapLogic states that you can use either ETL or ELT data load patterns and, after enriching data in a cloud data warehouse, connect to operational applications with reverse ETL. SnapLogic describes its product as an agile, cloud-native iPaaS for data platforms. Separate activation tools send warehouse data to business applications. Hightouch states that it integrates with 300+ tools spanning Advertising, Marketing Automation, CRMs, Customer Success, and more.
| Vendor | Approach | Where it fits |
|---|---|---|
| AWS Glue | Cloud provider services | Serverless ETL, ELT and streaming jobs on AWS, billed by DPU-hour |
| Microsoft Fabric / Azure Data Factory | Cloud provider services | Data Factory in Fabric, the next generation of Azure Data Factory, on Fabric capacity |
| Informatica / Salesforce | Enterprise data integration suites | ELT/ETL, replication and CDC in IDMC; Informatica is now owned by Salesforce |
| Qlik Talend | Enterprise data integration suites | Qlik Talend Cloud and client-managed Qlik Data Integration; Stitch is now part of Qlik |
| Fivetran + dbt Labs | Managed ELT and transformation | Managed connectors from 750+ sources, dbt for SQL transformation, and activation back to applications |
| Airbyte | Open-source ELT | Connectors for 700+ sources, self-managed, hybrid or in Airbyte Cloud |
| Matillion | Visual pipeline builders | Maia Foundation, formerly Data Productivity Cloud, for Snowflake, Databricks, BigQuery and Redshift |
| Apache Airflow / Astronomer | Workflow orchestration | Open-source orchestration, run yourself, as Astro, or as a cloud provider service |
Eight vendors are named across six approaches, and two approaches hold two vendors each. The approaches were set before any vendor was selected, and a place in this table is not for sale. The field is wider, including application integration vendors with data pipelines and smaller managed ELT services, and any data integration vendor can make its own case in the Spotlight below.
How much should you budget for ETL & ELT Data Integration Tools?
Data integration is priced on what moves, what runs, or what you reserve. Each model rewards a different workload, so price your own sources and volumes before comparing quotes.
Row-based pricing counts changed records. Fivetran defines MAR as the number of distinct rows synced from the source system to your destination system in a given calendar month, determined by tracking distinct primary keys. Under that definition the bill follows how many distinct rows change each month, not how often each one changes. Hevo Data counts each record that is inserted, updated or deleted in the destination as one event.
Volume pricing counts gigabytes. Estuary states that its Cloud plan costs $0.50 per GB of data plus $100 per connector instance per month, dropping to $50 for each instance after your sixth. A per-gigabyte price is easy to estimate from your data size. Add the fee for each source and destination you run.
Compute pricing counts the time jobs run. AWS charges an hourly rate for AWS Glue jobs based on the number of Data Processing Units (DPUs) used, where a single DPU provides 4 vCPU and 16 GB of memory, billed in increments of 1 second. Matillion prices on credits that are consumed based on actual task hours for orchestration, transformation, and streaming work. Airbyte states that its Pro and Enterprise Flex plans use capacity-based pricing built on Data Workers, the dedicated compute units that power your pipelines. Airbyte offers Standard, Plus, Pro, and Enterprise Flex, and states that Standard is fully managed volume-based cloud. Long jobs on small data are the expensive case, so check job size and schedule before the first invoice.
Platform pricing draws several services from one reserved pool. Microsoft describes Microsoft Fabric Capacity as a shared pool of capacity that powers all capabilities in Microsoft Fabric, from data modeling and data warehousing to business intelligence and AI experiences. Microsoft states that for each intelligent optimization throughput resource usage in a pipeline execution, you consume 1.5 CU hours for data movement Copy activities, and that each orchestration activity run consumes 0.0056 CU hours. Informatica states that Informatica Processing Units (IPUs) give customers access to all eligible cloud services listed in its Cloud and Product Description Schedule. Qlik states that Qlik Talend Cloud comes in four editions, and that customers subscribe to a certain amount of capacity of usage for the chosen edition, measured on a combination of data volume moved, number of job executions and execution duration. A pool simplifies purchasing, and integration then competes with every other workload for the same capacity.
Orchestration is priced separately. Astronomer states that workers on Astro automatically scale to zero when idle, with seven worker sizes starting at $0.13/hr. An open-source scheduler you host yourself has no license fee. Price the engineers who keep it running.
| Meter | You are charged for | Grows with | Where it goes wrong |
|---|---|---|---|
| Rows changed | Distinct rows inserted, updated or deleted in a month | Source activity and number of connections | High-churn tables and duplicate connections to one source |
| Data volume | Gigabytes captured and delivered | Data size and number of destinations | One source delivered to many destinations |
| Compute time | Processing units or worker hours while jobs run | Job length, frequency and size | Oversized jobs and frequent schedules on small data |
| Platform capacity | A reserved pool drawn on by several services | Every workload on the platform | Integration competing with analytics for the same capacity |
| Connector fees | Each active source or destination | Number of sources | Many small sources, each with its own fee |
How long does implementation take for ETL & ELT Data Integration Tools?
A data integration rollout replaces pipelines one source at a time. Old and new pipelines run side by side until their numbers match.
List every source, table, volume, change rate and schedule, and the reports and models that depend on each. Mark the sources with deletes, large tables and frequent schema changes.
Connect the three hardest sources on each shortlisted tool. Measure freshness, load on the source, failure recovery and the bill at production volume.
Move sources in waves. Reconcile row counts and totals against the old pipeline before switching each report, and keep a rollback path until the totals agree.
Alert on freshness per table, review cost per connection each month, and retire old jobs and credentials once nothing reads from them.
What should you ask vendors about ETL & ELT Data Integration Tools?
Ask each vendor to answer in writing and to demonstrate on your own sources and volumes.
-
Will transformation run inside a cloud warehouse or lakehouse?Yes Shortlist ELT tools and price the warehouse compute for transformation alongside the tool.No Shortlist ETL engines and integration suites that transform in flight, and size them for your heaviest job.
-
Do you need inserts, updates and deletes from operational databases within minutes?Yes Test log-based CDC on each database you run, and measure the load it puts on the source.No Scheduled batch loads may be enough. Compare connector coverage and price.
-
Must data stay inside your own network?Yes Compare self-managed and hybrid deployments, and count the infrastructure you will run.No Compare managed services on connector maintenance, freshness and the bill at your volume.