Well, a lot of disk writing even with adding the 'CLEAR_Selection' after the last time step. So more is going on behind the scenes that meets the eye. My cache bloomed. I ran it with ArcGIS Pro open and closed. and I got results similar to yours
Here is a query I did on a 2K table which results in 492 records being selected
import arcpy
import time
def test():
""" """
gdb = r"C:\Path_to_my_gdb\array_tools.gdb"
start = time.time()
lyr = gdb + r"\pnts_2000"
qry = "Text01 = 'C'"
r = arcpy.management.SelectLayerByAttribute(lyr, "NEW_SELECTION", qry, None)
stop = time.time()
arcpy.management.SelectLayerByAttribute(lyr, "CLEAR_SELECTION")
return r, (stop - start)
r, t = test()
t => 0.238 s
r[0] => <arcpy._mp.Layer at 0x226a6ec46a0> The result object, layer
r[1] => '492' The result count (which is correct)
running it 10 times yielded
[test() for i in range(10)][(<Result 'pnts_2000_Layer20'>, 0.25133728981018066),
(<Result 'pnts_2000_Layer26'>, 0.20045042037963867),
(<Result 'pnts_2000_Layer32'>, 0.24329805374145508),
(<Result 'pnts_2000_Layer38'>, 0.22243928909301758),
(<Result 'pnts_2000_Layer44'>, 0.23232150077819824),
(<Result 'pnts_2000_Layer50'>, 0.2583131790161133),
(<Result 'pnts_2000_Layer56'>, 0.25226902961730957),
(<Result 'pnts_2000_Layer62'>, 0.3051586151123047),
(<Result 'pnts_2000_Layer68'>, 0.282210111618042),
(<Result 'pnts_2000_Layer74'>, 0.2693297863006592)]
about the same... this did scale quite large when it got over 50 into 100 range.
Couldn't be that slow... so I cranked up my usual method of doing queries, using the same file read from disk.
The actual conversion to a numpy array took 0.0075 seconds with a whole 17.5 microseconds per query for 7 runs of 100,000 loops.
%timeit arcpy.da.TableToNumPyArray(lyr, "Text01")
7.5 ms ± 132 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
%timeit (a['Text01'] == 'C')
17.5 µs ± 63.2 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
If you are really performing repetitive queries or even different queries, you might want to consider this alternative.
To that end, I determined all the unique cases in that table's field and ran a query on each of them and timed the whole lot.
np.unique(a)
array([('A',), ('Aa',), ('Aaa',), ('B',), ('Bbbbb',), ('C',), ('Ccccc',),
('D',), ('Dddd',), ('None',)], dtype=[('Text01', '<U5')])
%timeit (a['Text_01' == i] for i in np.unique(a))
537 µs ± 15.5 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
Scaling quite nicely for 10 queries for the 10 cases, run 7 times for 1000 loops.
Options abound since NumPy and esri products mesh quite nicely and offer advantages to the standard interface for large data analysis and/or just about any task.
