Firestore: using ArrayUnion in 'update()' or 'set(merge=true)' on nested data removes all previous data

[A-Iskakov]

Environment details

1. Cloud Firestore
2. Ubuntu 18.04
3. Python 3.7
4. google-cloud-firestore Version: 0.31.0

Steps to reproduce

1. Use ArrayUnion([1,2,3]) on updating nested data in Document reference (see code example bellow)

Code example

from google.cloud import firestore
from google.cloud.firestore_v1beta1 import ArrayUnion

# creating reference
db = firestore.Client()
doc_ref = db.collection(u'test_collection').document(u'test_document')

# creating initial data with nested tree_1
doc_ref.set(
            {
                u'forest': {
                    u'tree_1': u'oak'
                }
            }
        )

# getting document after initial data
print('Here is initial data with 1 nested tree', doc_ref.get().to_dict())

# adding tree_2 to nested data (tree_1 remains in nested data)
doc_ref.set(
            {
                u'forest': {
                    u'tree_2': u'pine'
                }
            },
            merge=True
        )

print('Here is 2 trees in nested data after using set(merge=true)', doc_ref.get().to_dict())

# this code removes other nested data in document
# including tree_1 and tree_2, even if merge=true  is used
doc_ref.set(
            {
                u'forest': {
                    u'tree_3': ArrayUnion([u'spruce'])
                }
            },
            merge=True
        )

print('after using ArrayUnion, all previous nested data is removed', doc_ref.get().to_dict())

Stack trace

Here is initial data with 1 nested tree {'forest': {'tree_1': 'oak'}}
Here is 1 initial tree and 1 additional tree after set(merge=true) {'forest': {'tree_1': 'oak', 'tree_2': 'pine'}}
after using set(merge=true) with ArrayUnion all nested data is removed {'forest': {'tree_3': ['spruce']}}

Many Thanks!

[tseaver]

@A-Iskakov Thanks for the report. I believe this behavior is due to the same underlying cause as googleapis/google-cloud-python#7215: the code leaves behind the parent mapping in the update_mask when moving the transform (ArrayUnion in your case, SERVER_TIMESTAMP in googleapis/google-cloud-python#7215) to the second write of the commit.

[A-Iskakov]

@tseaver Hello!
Thank you for reply!

I believe this behavior is due to the same underlying cause as googleapis/google-cloud-python#7215

I believe not as I have tried to replace ArrayUnion with SERVER_TIMESTAMP in the code to reproduce current issue, it works fine, nothing is deleted.
So the possible cause of problem in googleapis/google-cloud-python#7215 might be use of dots while addressing nested objects like 'foo.now'

[A-Iskakov]

@tseaver hello
I double checked the issue with SERVER_TIMESTAMP:
when SERVER_TIMESTAMP is only data to be sent with set(dict, merge=True) function, all original data is also erased.
if you add any other data to nested dictionary among with SERVER_TIMESTAMP, function works correct

[HemangChothani]

@A-Iskakov hello
I have checked issue it works corrrect, you just have to use ArrayUnion([u'spruce'].values, else it throws a TypeError.

Code Example

from google.cloud import firestore
from google.cloud.firestore_v1beta1 import ArrayUnion

db = firestore.Client()
doc_ref = db.collection(u'test_collection').document(u'test_document')

# creating initial data with nested tree_1
doc_ref.set(
            {
                u'forest': {
                    u'tree_1': u'oak'
                }
            }
        )
print('Here is initial data with 1 nested tree', doc_ref.get().to_dict())
doc_ref.set(
            {
                u'forest': {
                    u'tree_3': ArrayUnion([u'spruce']).values
                }
            },
            merge=True
        )

print('after using ArrayUnion' , doc_ref.get().to_dict())

Output:

Here is initial data with 1 nested tree {'forest': {'tree_1': 'oak'}}
after using ArrayUnion {'forest': {'tree_1': 'oak'},{'tree_3': ['spruce']}}

[tseaver]

I can confirm that the issue remains even after the merge of googleapis/google-cloud-python#8701

Python 3.7.1 (default, Oct 22 2018, 11:31:38) 
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from google.oauth2 import service_account
>>> from google.cloud import firestore
>>> creds = service_account.Credentials.from_service_account_file('/path/to/creds.json')
>>> client = firestore.Client(credentials=creds, project=creds.project_id)
>>> doc_ref = client.document('gcp-7523', 'test-document')
>>> doc_ref.get().to_dict()
>>> doc_ref.set({'forest': {'tree_1': 'oak'}})
update_time {
  seconds: 1563560252
  nanos: 187033000
}

>>> doc_ref.get().to_dict()
{'forest': {'tree_1': 'oak'}}
>>> doc_ref.set({'forest': {'tree_2': 'pine'}}, merge=True)
update_time {
  seconds: 1563560269
  nanos: 162782000
}

>>> doc_ref.get().to_dict()
{'forest': {'tree_1': 'oak', 'tree_2': 'pine'}}
>>> doc_ref.set({'forest': {'tree_3': firestore.ArrayUnion(['spruce'])}}, merge=True)
update_time {
  seconds: 1563560280
  nanos: 164830000
}

>>> doc_ref.get().to_dict()
{'forest': {'tree_3': ['spruce']}}

@jadekler, @schmidt-sebastian ISTM that the behavior here (set-with-merge of data with nested transforms wipes out other nested keys) is mandated by https://github.com/googleapis/conformance-tests/blob/master/firestore/v1/set-st-merge-nonleaf-alone.json. Can you confirm?

[jeanbza]

cc @BenWhitehead who has taken over those tests / firestore.

[tseaver]

@jadekler Thanks, sorry for the noise.

This is my reproducer system test for this issue:

def test_repro_7523(client, cleanup):

    doc_ref = client.document("gcp-7523", "test-document")
    cleanup(doc_ref.delete)
    doc_ref.delete()
    tree_1 = {"forest": {"tree-1": "oak"}}
    tree_2 = {"forest": {"tree-2": "pine"}}
    tree_3 = {"forest": {"tree-3": firestore.ArrayUnion(["spruce"])}}

    doc_ref.set(tree_1)
    expected = tree_1.copy()
    assert doc_ref.get().to_dict() == expected

    doc_ref.set(tree_2, merge=True)
    expected["forest"]["tree-2"] = tree_2["forest"]["tree-2"]
    assert doc_ref.get().to_dict() == expected

    doc_ref.set(tree_3, merge=True)
    expected["forest"]["tree-3"] = "spruce"
    assert doc_ref.get().to_dict() == expected

[BenWhitehead]

After chatting with @schmidt-sebastian this issues does appear to be a bug, and the merge with transform should work with nested fields.

We identified this conformance test set: ServerTimestamp alone with MergeAll as similar to this issue, however this issue is updating in a nested map whereas the test is not.

I'm going to look into adding a new test for this case and we can ensure it's working for all the clients.

Internal bug for new test b/138130675

[tseaver]

@BenWhitehead OK, cool. Please do also consider the set-st-merge-nonleaf-alone test (whose description reads "set-merge: non-leaf merge field with ServerTimestamp alone").

[tseaver]

@BenWhitehead Any update on this issue? I don't see any new commits to the firestore conformance tests.

[amitjain-v]

Is there any known workaround for this issue ?