Controlling Azure-SSIS Integration Runtime Cost: Schedules, Node Size, and Licence Mode

The first Azure invoice after a lift-and-shift is where most SSIS migrations get their reputation. The packages work, reconciliation passed, and the line item is three or four times what anyone put in the business case.

The cause is almost never the packages. It is that the Azure-SSIS integration runtime bills by the node-hour from the moment it starts until someone stops it, and nothing stops it by default. An estate that runs for ninety minutes a night on-premises can bill for twenty-four hours a day in Azure and produce identical output.

Four settings decide most of the bill. None of them are hard. They are just easy to leave at the wizard default.

1. Start and stop the runtime around the work

This is the big one, and it is usually worth more than the other three combined.

The Azure-SSIS IR has two meters: the compute nodes, and — if you used the wizard default — the SSISDB catalog database. Compute is charged per node, per hour, while the IR is in the Started state. A stopped IR charges nothing for compute. The catalog database keeps charging either way, which is a separate conversation about service tier.

So the question is simply: how many hours a day does the IR actually need to be up? For most mid-market estates the honest answer is two to four. Nightly loads finish before the business arrives, and nothing touches SSIS until the next night.

Azure Data Factory ships two pipelines for this. Create them from Manage → Integration runtimes → your IR → ... → Create start/stop schedule, which generates a pair of pipelines wrapping the StartAzureSSISIR and StopAzureSSISIR web activities, each on its own schedule trigger.

The better pattern, once you trust it, is not a stop trigger on the clock at all. Make the start the first activity of your nightly master pipeline and the stop the last one, so the runtime is up for exactly as long as the load takes:

Start IR  ->  Execute SSIS Package (master)  ->  Stop IR

Two cautions with that pattern. Set the Stop activity's dependency to Completed, not Succeeded, or a failed load leaves the runtime running all weekend. And allow for the start time: an Azure-SSIS IR takes roughly 20 to 30 minutes to come up, so the first package does not begin when the trigger fires. If a 06:00 SLA matters, start the IR at 05:00, not 05:55.

If you want a safety net rather than a schedule, add a second trigger that calls the stop pipeline every morning regardless. It costs nothing when the IR is already stopped.

2. Size the node to the package, not to the server it came from

Node size is the second meter. People size the IR against the old physical server — 16 cores because the SQL box had 16 cores — and then run packages that never used more than two.

SSIS data flows are mostly memory-bound, not core-bound. The practical sequence is:

  1. Start at a mid-range general-purpose size.
  2. Run the real nightly load.
  3. Look at the package execution times and the node metrics before changing anything.

If the data flows are spilling buffers, more memory helps and more cores do not. If packages are serialised behind each other, the fix is parallelism (below), not a bigger node. If the slow step is a SQL query on the source, the node size is irrelevant and you should be tuning the query.

Measure first. We do not guess at a size before a package has run on it, and neither should a vendor quoting you one.

3. Node count and maximum parallel executions

Node count multiplies the hourly rate directly: two nodes cost twice one node for the same wall-clock hour. That is worth it only if it shortens the window enough to matter, or if the estate genuinely needs the concurrency.

Before adding nodes, check MaxParallelExecutionsPerNode. It defaults conservatively, and on a node with spare cores raising it lets more packages run concurrently on hardware you are already paying for. One larger node at higher parallelism is often cheaper than two small ones, because you pay the startup window once.

The counter-argument: two nodes survive a node failure and one does not. For a nightly batch with a retry and a morning of slack, that is usually an acceptable risk. For a load that feeds a regulatory deadline, it is not. Decide it deliberately rather than by default.

4. Licence mode and Azure Hybrid Benefit

The Azure-SSIS IR is priced in two ways: licence included, or bring-your-own SQL Server licence under Azure Hybrid Benefit. If your organisation has SQL Server Enterprise or Standard licences with active Software Assurance — and most estates migrating off SQL Server 2016 do — the hybrid option removes the SQL licence component from the node rate.

It is a dropdown on the IR, and it is the only cost lever here that requires no engineering at all. It does require a straight answer from whoever owns your Microsoft agreement about which licences have Software Assurance and whether they are already allocated elsewhere. Ask early; that answer takes longer to get than the change takes to make.

The same question comes up for the catalog. If the SSISDB catalog sits on SQL Managed Instance, Hybrid Benefit applies there too, and the Managed Instance is usually the larger line item of the two.

The order to do this in

  1. Put a start/stop schedule on the runtime this week. It is the largest saving and the lowest risk.
  2. Check the licence mode. One dropdown, if the entitlement exists.
  3. Run a real nightly cycle, then look at the execution times before touching node size or count.
  4. Raise parallel executions per node before adding a node.

When the answer is to change nothing

If the estate is small — a handful of packages, a few minutes a night — the IR may be the wrong host entirely. Packages that are a source query, a Derived Column and a destination are a Copy activity in Data Factory, with no runtime to start, stop or size. We would rather tell you that than tune an invoice you did not need.

And if the packages still run fine on a server you already own, keeping them there is a legitimate answer. The deadline that moves most estates is SQL Server 2016 leaving extended support on 14 July 2026, not anything about SSIS itself.

If you are costing a move and want the node sizing based on what your packages actually do, tell us how many packages you have and what they run on — a senior engineer replies within a day. The sizing falls out of an estate assessment, and the start/stop and licence work is part of a lift-and-shift.