You are doing it again: exporting last week's orders from Shopify or your CRM, pasting them into a spreadsheet, fixing duplicate rows and date formats that never agree. If you’re wondering what is data warehouse.">ETL, the answer is that it automates this routine. It pulls data from your tools, cleans it according to rules you define once, and loads it into one place so the report builds itself. The ETL safety-mark and the accounting firm of the same name are not what this page is about.
What ETL stands for and what it actually does
ETL stands for extract, transform, load. A routine copies data from your various systems, applies cleaning rules to make it consistent, then stores it in a single destination like a data warehouse where reporting tools can read it.
The process dates from the 1970s, when it was introduced for integrating and loading data for analysis, but the underlying need has not changed: businesses have always wanted one trusted monthly picture built from scattered sources. What is new is that both the sources and the targets now live in the cloud. This makes the whole process cheaper to run and easier to automate than it used to be.
The steps named plainly
- Extract: Copy raw rows out of your systems (Shopify orders, Xero invoices) into a temporary holding area.
- Transform: Apply your business rules. Remove duplicates, standardise date formats, convert currencies if you trade across borders, and calculate the metrics you actually report on.
- Load: Write the cleaned rows into the destination. The first run is a full load; every run after that is incremental, moving only what has changed since last time.
The business purpose is reporting. If you produce a sales summary or an invoice reconciliation weekly, ETL is simply the plumbing that gets those numbers to you without your hands in them.
What happens inside each step, with a real example
Extract: getting data out of the right places
A small UK shop might pull from Shopify for orders and customers, Xero for invoices and payments, and a CRM for lead status. A connector tool handles the extraction: it reads whatever changed since the last run and drops those rows into a staging area. That buffer keeps messy, mixed-format data from landing straight in your reporting layer.
Transform: the fixes you do by hand today
This is where ETL saves the most time. The transformation rules mirror what you already do manually: removing duplicate customer entries when they come in from two sources, converting dates to one consistent format so "1 Oct 2026" and "10/2026" stop colliding. You might also calculate margin per product line or match invoice numbers to bank payment references where clients have not typed the reference correctly.
Write these rules once; every subsequent run applies them automatically.
Load: getting data into one place you trust
The destination is normally a data warehouse. The first load pulls everything in, which can be slow if you have years of history. After that, each scheduled run does an incremental load, storing only new or changed rows since the last extract date. That keeps the process cheap and fast while still giving you a complete, up-to-date table to query.
A worked example
Picture a small agency with five clients. Every month someone manually combines invoice data from Xero, time entries from an internal tool, and project status from a CRM into a spreadsheet so they can see which projects are profitable. With ETL, the connector pulls new invoices, time rows and status changes each week. The transform step strips out test entries, standardises client names (because "Acme Ltd" and "acme-ltd" are the same account), and calculates hours billed versus hours budgeted per project.
The load lands one clean table in BigQuery that a Metabase dashboard reads. The Monday report is there before you open your laptop.
ETL versus ELT, and which one you will actually use
The difference comes down to when the transformation happens. In classic ETL, data is cleaned before it is loaded into the destination. In ELT (extract, load, transform), raw data goes straight in and the cleaning and modelling happen inside the warehouse after loading. That is the only structural difference; everything else is the same pipeline doing the same job.
ELT became the default with cloud warehouses because the compute power to do heavy transformations lives right where your data sits. What we see is that most small businesses that use a tool like Airbyte or Fivetran are effectively running ELT without thinking about it: raw rows land in BigQuery first, and the business rules are applied inside the warehouse.
What changes for you as the owner
ETL carries higher upfront cost because you need infrastructure and engineering to do the transformation before data arrives. It also needs re-engineering when source systems change structure.
ELT is simpler to set up, but if your data volumes grow unexpectedly, query and storage costs can surprise you on the bill. We build our client warehouses as ELT: raw data lands first, then we model it inside BigQuery. Use whichever term your vendor uses; the outcome for you is the same clean table behind your dashboard.
Signs a small business actually needs this
You do not need ETL if you have one source of truth and one person who can read it. You start needing it when the Monday report draws on three or more systems and someone has to reconcile them by hand every week. A few situations that say yes:
- You match 40 to 50 invoices against bank payments manually each month, and clients frequently omit reference numbers.
- Your lead tracking lives in sticky notes plus a spreadsheet, and you have missed callbacks because the system is too fragile to trust.
- The business runs on one person’s memory of what happened last quarter, and that person is also your only salesperson.
If you only need a monthly export from one tool into a pivot table, or if two of your apps already integrate directly, a connector or even a scheduled CSV pull may be enough. ETL becomes necessary when numbers from several systems must agree on the same day without a human in the loop. And because it is the plumbing behind a data warehouse, if you do not yet need one of those, you probably do not need this either.
Does this apply to you?
Probably not yet
If none or only one of these sounds familiar, a well-structured spreadsheet is still the right tool.
What it costs and how long it takes to set up
Connector tools
Airbyte Standard starts at $20 per month with about 1 GB of syncs included, priced in US dollars at the time of writing. Airbyte Core is free to self-host if you have a server to run it on, or you can start free for 14 days with 400 credits and no credit card required.
Fivetran offers a free tier covering up to 500,000 monthly active rows across connections; beyond that the pricing is usage-based. Their own example shows a Standard plan estimate of $511.15 per month for four connectors, which illustrates how quickly costs scale if you connect many sources or move large volumes.
Warehouse storage
BigQuery charges by storage volume and query usage. At the time of writing, storing 100 GiB for half a month in us-central1 costs $1.15 USD, and 1 TiB for a full month costs $23.552 USD. Data that has not been modified or deleted within 90 consecutive days drops to the long-term storage rate, so historical data gets cheaper over time.
Time to build
Connecting your first source and getting rows landing in BigQuery can take an afternoon if the connector exists for your tool. Getting a trustworthy weekly report that multiple people rely on is typically a matter of weeks: most of that time goes into agreeing what "revenue" or "active lead" means before you write a single transformation rule.
Ongoing, budget attention when a source changes its export format or renames a field; that is the hidden cost nobody quotes you upfront. We have built small-business warehouses on this stack: Airbyte pulling sources, BigQuery as the destination, Metabase reading the clean tables. The build itself is straightforward once definitions are agreed; the long tail of maintenance is where most of your hours go.
In practice · Case study
Building a Data Warehouse
Turning five disconnected platforms into one source of truth.
Read the case studyWhat goes wrong and who fixes it
The most common failure mode is a source changing without telling you. A CRM renames a column, an e-commerce platform adds a new field to the order object, or an accounting export shifts its date format. The pipeline still runs, but now it loads garbage or drops rows silently. You will not see an error; you will just notice the weekly report is wrong until someone spots it.
Incremental loads make this worse because they rely on a reliable "what changed" marker, usually a timestamp or ID. If your source does not update that marker correctly, you either miss new data or load duplicates. Both problems are invisible until the numbers stop matching reality.
Even experienced data engineers complain about how poorly documented basic ETL tasks are; beginner-friendly guides that walk through real tools from the ground up are scarce. That is a problem with the ecosystem, not with your skill level.
The practical question you need to answer before anything else: who looks at this on a Monday morning when something breaks? Is it you, a freelancer on retainer, or us? Name that person in writing before you start building, because an unowned pipeline is just another spreadsheet nobody trusts.
What to do next
Open a blank document and list every system that feeds your Monday or monthly report. Next to each one, write the single number that is most often wrong when you pull it together by hand. If any of those numbers comes from two different systems, you have a business case for at least trying an automated pipeline.
Start with one source. Pick the connector tool whose free tier fits your volume (Airbyte Core if you can self-host, Fivetran Free if you cannot), and get that single stream landing in a warehouse or even a well-structured Google Sheet so you can see the data moving without touching it manually. You will know within a week whether the rest is worth building.
If after that trial you still want someone to design the full pipeline, agree the metric definitions with your team and own the maintenance. That way a broken connector on a Tuesday does not become a month of wrong numbers, and it is what our data and reporting service covers. Bring the list you just wrote down and we will discuss in a short call whether a pipeline beats doing nothing.
Key terms
- ETL
- Extract, transform, load – a three-step process that moves data from source systems, cleans it, and stores it in a destination such as a data warehouse.
- ELT
- Extract, load, transform – the same pipeline but with transformation happening inside the warehouse after the raw data has been loaded.
- data warehouse
- A central repository where cleaned data is stored for reporting and analysis; BigQuery is a common example for small businesses.
- staging area
- A temporary storage location that holds raw rows after extraction, allowing them to be inspected and cleaned before loading into the warehouse.
- connector
- A pre-built integration that reads data from a SaaS tool (e.g., Shopify, Xero) and writes it to a staging area or directly to a warehouse.
- incremental load
- A load that only transfers rows that have changed since the previous run, keeping the pipeline fast and cheap.
Frequently asked questions
What does ETL stand for?
Is SQL an ETL tool?
What is the difference between ETL and ELT?
What is an ETL pipeline?
Is Python used for ETL?
Do I need a data warehouse to use ETL?
Sources
- What is ETL (extract, transform, load)? | IBM, IBM (accessed 26 Oct 2026)
- What is ETL? | AWS, Amazon Web Services (accessed 26 Oct 2026)
- What is ETL? | Google Cloud, Google Cloud (accessed 26 Oct 2026)
- Extract, transform, load - Wikipedia, Wikipedia (accessed 26 Oct 2026)
- Airbyte pricing, Airbyte (accessed 26 Oct 2026)
- Fivetran pricing, Free plan, Fivetran (accessed 26 Oct 2026)
- BigQuery pricing, Google Cloud (accessed 24 Oct 2026)