Macro's are compile time, there is no runtime codegen.

The problem was overly-frequent inlining generating enormous expressions, causing a lot JIT phase and slow execution.

Thank you for the clarification. If I understand correctly, these large expressions are created at compile-time, but the impact isn't felt until JIT occurs in the runtime environment. In that scenario, shouldn't the JIT just run once at startup, though? I'm still not quite understanding how JIT can take so much time in a production environment.

▲

hunterpayne a day ago | parent [-]

Because the jit will let the unoptimized code run a few (hundred) times to take measurements to know what needs to be optimized and how it needs to be optimized. This is a good solution and makes hotspot very effective. The problem is that it happens randomly a few minutes/seconds into the operation of the service. So you randomly have a big pause with the performance hit everytime you run the service. The upside is that this only happens once. But you have to plan for a big performance hit to requests which are unlucky enough to be called at the wrong time.

▲

pretzellogician a day ago | parent | next [-]

And this can generally be avoided as well, by doing "warmup" when starting your service (effectively, mock some calls), but before accepting requests.

	▲	hunterpayne a day ago \| parent [-]
		Of course, but then you have to actually do this. It is just another complexity to add. Also, I was answering a question about the hows and whys of the jit. I wasn't saying it was impossible to work around.

▲

munchler a day ago | parent | prev [-]

Ah, that’s interesting. I wasn’t aware that JIT-ing will do that sort of performance analysis first. Thank you for the explanation.

▲

gavinray a day ago | parent | prev [-]

That's not true, Spark's entire query engine relies on use of runtime codegen via macros/quasi quotes

Look up the architecture of Catalyst + Tungsten

https://www.databricks.com/glossary/catalyst-optimizer

	▲	lmm a day ago \| parent [-]
		Catalyst uses runtime codegen, sure, but the OP wasn't using that.