Code Translation
Disclaimer: I work for Next Pathway, which performs code translation as part of a wider offering to migrate customers from legacy environments to the cloud.
I speak to our product teams regularly, I work with our delivery teams often, although they are part of a well oiled machine I see the complexity that they are managing in performing migrations.
And with that in mind I’m seeing a large number of global services organisations and smaller more product focused companies starting to offer migration capabilities based on Gen AI (or more precisely LLMs) and I wonder can this be true?
The first thing that we need to do in any analytics migration project is define what we have today, and what we will migrate. We need to answer questions such as:
- What are the interdependences?
- I have pipelines that include a source system, an ETL tool, a data platform, more ETL processing, an output schema, a semantic layer and a bunch of consumption tools - What can we immediately determine does not need to be migrated?
- How would I figure this out? I don’t want to spend time and money migrating rubbish - Who decides what needs to be migrated first?
- In any large system, there will be multiple stakeholders — who is shouting loudest? If I migrate this set of artefacts first does it accelerate the migration of subsequent artefacts?
….and I think how does the LLM help here? Can I feed it scheduling, orchestration, SQL, data platform specific functions, an export of a semantic layer, BI report definitions and analytical models and ask it to create all of the dependencies and expect an accurate answer?
My suggestion is probably ‘No’.
Could I make this work if I had a lot of expertise in the legacy and could manage the code in the same way we need to manage data for LLMs….
Get Andrew Mott’s stories in your inbox
Join Medium for free to get updates from this writer.
The second thing that we need to do in any migration is translate the code from the legacy systems to the target systems. This is the easy part I’m not an expert and I can get an LLMs to generate all manner of text and code, but we still need to answer questions such as:
- How can I determine what code from which legacy system needs to be converted to which target system?
- What is scheduling?, what is orchestration?, what is extraction code? What is ingestion code? - How do we translate code accurately?
- How do we reduce hallucinations? What if the LLM generates ELT code and expects the BI tools to execute it? - How can we ensure that the generated code is maintainable and manageable?
….and I think can an LLM help here? Of course, like I said I can get one to generate all manner of nonsense, but will the code be useful?
My suggestion is probably ‘No’.
The third part of a migration is testing, testing that the code works, testing the target systems against the legacy systems. To do this we need to think about some more questions:
- What do we test?
- How do we test?
- Where do we test?
- What happens if the code doesn’t work?
- How do I determine why a number is incorrect?
- How do I determine what the actual number should be?
And can an LLM help here? That is a good question, could we generate test cases with an LLM? Yes
Could we use the LLM to test if test cases are valid before we apply them to our target environment? Yes
But this seems like we might need to recursively test the test cases forever before we can apply them to test the thing we want to test.
It feels like an automated deterministic approach might actually be what we want here…. I guess we might want to think a bit more before believing that an LLM will solve all of our migration questions….
But perhaps I’m biased….
