[Data] RandomAccessDataset.multiget return unexpected values for missing keys.

[sunyakun]

What happened + What you expected to happen

the ray.data.RandomAccessDataset.multiget expected return a None for missing records, in fact, I got an unexpected value for the missing key.

I find this PR update the _RandomAccessWorker.multiget: #24825, and it use the np.searchsorted to speed up the multiget, but the np.searchsorted will return the insertion points for missing records and it use the search result directly to get the row from the block without test col[i] == key, just like the code here:

ray/python/ray/data/random_access_dataset.py

Lines 266 to 269 in d8c7234

	i = bisect.bisect_left(column, x)
	if i != len(column) and column[i] == x:
	return i
	return None

Versions / Dependencies

Ray: latest master
Python: 3.9.2
OS: linux

Reproduction script

import ray
import ray.data

kv_store = ray.data.from_items(
    [i for i in range(0, 1000, 2)]
).repartition(5).to_random_access_dataset(key="item", num_workers=1)

print(ray.get(kv_store.get_async(1)), ray.get(kv_store.get_async(901)))
# output: None None

print(kv_store.multiget([1, 901]))
# output: [{'item': 2}, {'item': 902}]

Issue Severity

None

[tespent]

I can reproduce this problem and I created a pull request #44769 trying to fix this.