To have more statistics on this power drop at DF issue, I fitted the decay curves of photodiode signals.
In June, we had the time constant of ~90 sec, but from July, we have the different time constant of ~50 sec.
The power drop in June was ~10 %, and in June was ~30% (as already reported in logbook #42058).
Time constant being different adds another mystery.....


Method:
1. Get the data of META_ITF_State, B2_DC, B4_DC, B5_DC, B7_DC, B8_DC, B4_12MHz, and B4_112MHz. Extract the lock which reached Low Noise 3, and plotted the data from -100 sec to +1000sec from the point we reached LOCKED_PRITF_DF.

2. Fitted decay curves with a*exp(x/tau)+c for each lock. For DC signals, I used the data from LOCKED_PRITF_DF for 400 sec. For B4_12MHz, I used the data from 50 sec after LOCKED_PRITF_DF for 300 sec since it fluctuates too much at the beginning of DF.

3. Plotted tau and the power drop ratio r=a/c for each lock, and calculated the median (not average since B4_12MHz fluctuates a lot and fitting is not reliable in many cases) and standard deviation.


Result:
The time series data of 34 locks between June 21 to 25 is shown in DFpowerdrop_LOCKED_PRITF_DF_1213574418_432000.png.
The time series data of 13 locks between June 23 to 27 is shown in DFpowerdrop_LOCKED_PRITF_DF_1213574418_432000.png.
Fitted curves are also plotted as dashed black lines. As you can see, the power drop is consistent between locks (semilar plot for B7_DC also in logbook #42225).
In June from June 21, 6 MHz was not decreased in DF (logbook #41932), so we can clearly see that B4 12 MHz also drops (as also reported in logbook #42011).
Note that B4 112 MHz do not decay at this time scale, but it also decays at longer the time constant of 1.5e3 sec (see logbook #41817).
In July, we decreased 6MHz in DF (logbook #42162), so it is hard to tell what's happening for 6 MHz sidebands.

Summary of fitting results are shown below.

From June 21 for 5 days, median of 34 locks (DFpowerdrop_LOCKED_PRITF_DF_1213574418_432000_fitresult.png)
LSC_B2_DC_mean tau=80.17 +/- 60.95 sec
LSC_B2_DC_mean r=35.02 +/- 6.75 %
LSC_B4_DC_mean tau=90.79 +/- 17.83 sec
LSC_B4_DC_mean r=9.79 +/- 0.99 %
LSC_B5_DC_mean tau=87.60 +/- 18.27 sec
LSC_B5_DC_mean r=9.95 +/- 0.90 %
LSC_B7_DC_mean tau=92.40 +/- 18.69 sec
LSC_B7_DC_mean r=9.69 +/- 0.89 %
LSC_B8_DC_mean tau=91.49 +/- 24.35 sec
LSC_B8_DC_mean r=9.48 +/- 0.97 %
LSC_B4_12MHz_MAG_mean tau=108.22 +/- 260.09 sec
LSC_B4_12MHz_MAG_mean r=39.68 +/- 152.40 %

From July for 5 days,  median of 13 locks (DFpowerdrop_LOCKED_PRITF_DF_1213574418_432000_fitresult.png)
LSC_B2_DC_mean tau=114.08 +/- 42.69 sec
LSC_B2_DC_mean r=24.61 +/- 6.15 %
LSC_B4_DC_mean tau=111.69 +/- 615.54 sec
LSC_B4_DC_mean r=7.28 +/- 4.59 %
LSC_B5_DC_mean tau=109.63 +/- 61.33 sec
LSC_B5_DC_mean r=7.45 +/- 2.11 %
LSC_B7_DC_mean tau=115.32 +/- 85.67 sec
LSC_B7_DC_mean r=6.73 +/- 2.53 %
LSC_B8_DC_mean tau=124.57 +/- 78.30 sec
LSC_B8_DC_mean r=7.10 +/- 2.42 %
LSC_B4_12MHz_MAG_mean tau=102.97 +/- 192.52 sec
LSC_B4_12MHz_MAG_mean r=33.93 +/- 106.92 %

As you can see, B4_DC, B5_DC, B7_DC and B8_DC have very similar time constants and power drop, which indicates power recycling gain decay, something in CITF. You can also see that B7_DC and B8_DC divided by B5_DC stays roughly constant from the time series plot. This means that finnesse of the arms stay roughly constant.
The time constant in June was ~10%, but it increased to ~30% in July. More surprisingly, time constant in June was ~90 sec, but it changed to ~50 sec in July.
B2_DC has slightly different time constant, and have more power drop.
B4_12MHz fluctuates a lot, and it is hard to tell, but it has similar time constant (~100 sec) with B4_DC, and more drop (~30%) compared with B4_DC.

B4_12MHz decaying faster than B4_112MHz indicate some alignment drift or abbreation drift at O(100) sec time scale, since 6 MHz is more sensitive to these effects.


Next:
 - Try not decreasing 6 MHz as much to see the behaviour of 6 MHz in DF, if possible.
 - Look for the mirror temperature drift (but it might be hard since line tracker (logbook #41883) demodulates B1, not B1p; B1 is not available in LOCKED_PRITF_DF and B1p is noisier).