GDALOpenShared does not properly handle open options

test-open-shared.py

"""
GDALOpenShared with open options.

This script demonstrates a race condition whereby
the first dataset opened in shared mode will take precedent
over subsequent datasets that are with different open options.

$ python test-open-order.py
2015 mean blue value: 77.47204
2017 mean blue value: 77.47204
2019 mean blue value: 77.47204

$ python test-open-order.py
2019 mean blue value: 82.556031
2017 mean blue value: 82.556031
2015 mean blue value: 82.556031
"""

import os
import random

from osgeo import gdal

GDAL_OF_SHARED = int("0x20", 16)
OPEN_FLAGS = GDAL_OF_SHARED | gdal.GA_ReadOnly

# Read a 1000x1000 chunk of a tileDB array with data at multiple timestamps
path = os.path.abspath("../naip/naip-combined")
x = y = 1000
cols = rows = 1000

years = [2019, 2017, 2015]
random.shuffle(years)

for year in years:
    src = gdal.OpenEx(path, OPEN_FLAGS, open_options=[f"TILEDB_TIMESTAMP={year}"])
    band = src.GetRasterBand(3)
    arr = band.ReadAsArray(x, y, cols, rows)
    print(f"{year} mean blue value: {arr.mean()}")

In a distributed/multi-process scenario serving web map tiles, this means your server nodes could be initialized with a different timestamp depending on the order and how the requests are routed. This makes for a fun, non-deterministic mixture of timestamps