I’m using the CosmosDB on Azure with MongoDbApi. In my Flask service I use pymongo.
My collection consists of a documents similar to this:

{
    "file_name": "foo",
    "date": "2024-05-07",
    "count": 1,
    "deliveries": [
        {
            ...,
            "events": [
                ...
            ],
            "type": "NG"
        },
        {
            ...,
            "events": [
                ...
            ],
            "type": "Legacy"
        }
    ],
    "snapshot_timestamp": 1234567
}

I have a set of functions (basically simple queries) that update the file. An example one would be something like this:

collection.update_one(
    {
        "file_name": ...,
        "snapshot_timestamp": ...,
    },
    {
        "$set": {
            "deliveries.$[outer].type": "NG",
            "deliveries.$[outer].foo": "bar",
            "count": 5
        },
        "$push": {
            "deliveries.$[outer].some_array": "some_value"
        }
    },
    upsert=True,
    array_filters=[
        {"outer.x": "y"}
    ]
)```

Sometimes I see a corrupted document that looks like this:
``` lang-json
{
    "file_name": "foo",
    "date": "2024-05-07",
    "count": 1,
    "deliveries": {
        "$[outer]": {
            ...,
            "events": [
                ...
            ],
            "type": "NG"
        },
    },
    "snapshot_timestamp": 1234567
}

I’m really trying to find the source of this, but there is no update operation where I use a single "$set": {"delivery.$[outer]"}. This happens exteremely rarely. Unfortunately I can’t find any type of trace or path that causes this. Is there any way where the update() operation can cause this behavior?