Could you please tell me how the sky fiber positions were selected?

Sky fibers used for the subtraction? I didn't specify them so config.selectSky.fiberFilter='ALL'.

I thought that this was a "proper" extra-Galactic field from HSC, so shouldn't we have SKY objects? Will we have SKY objects in the next engineering runs?

I don't think that this can explain the problem, though, as we are over subtracting the OH.

Murata-san's sky object DB was used and we have 500 SKY fibers (in entire focal plane not SM1&3) in this field.

Ah, OK. So they were specified in the pfsDesign, which is all that matters.

You shouldn't ever need to config.selectSky.fiberFilter except for interactive fiddling (a better solution is to write a pfsConfig to the rerun).

FWIW, attached is the location of fibers with large residual (using an isolated sky line) and sky fibers on PFI plane. It looks that there is no strong (anti-) correlation?

This is due to fibers being in the penumbra of nearby black spots. Based on a bunch of sky lines from two exposures (83149 and 83244), the throughput appears to be linear with distance from the black spot (slope of 1.232/mm) up to a distance of 1.108 mm.
I've added a correction for this effect in reduceExposure. The correction modifies the normalisation, so that the flux is still what is extracted from the image.
In order to keep track of the correction (and other reduction operations and quality measurements), I've added a new element to the spectra, currently called "notes". This is an addition to the datamodel, so needs approval via a DAMD ticket.
Using the fixed code, I get the following sky-subtracted merged spectrum for visit=83244 fiberId=478 (which observes objId=43158588759244917, which had the bad sky subtraction):

This includes the changes of DAMD-145 (adding reduction notes to spectral products) in order to record the values that were used.

Changes approved. Could not see any obvious issues with the proposed changes, including the new notes mechanism. Perhaps a slope to model the penumbra may be too simplistic for the future, but it works now and is a good start.

Need to check: it looks like I might be making the correction twice (once in extractSpectra, and once in reduceExposure).

I double-checked, and I am not making the correction twice. The second correction is in pfs.drp.stella.pipelines.extractSpectra (part of the Gen3 pipeline that replaces reduceExposure), not pfs.drp.stella.extractSpectra (the building block).

Merged to master.