Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,16 @@
_MUTATE_ROW_METHOD = "/google.bigtable.v2.Bigtable/MutateRow"
_READ_ROWS_METHOD = "/google.bigtable.v2.Bigtable/ReadRows"

# Inbound message cap on the UDS channel. gRPC defaults this to 4 MiB, a limit
# meant for a channel that crosses a network. This one doesn't leave the
# machine: the daemon has already paid the cost of pulling the response off the
# wire, and all the cap does on this hop is turn an otherwise-fine read into
# RESOURCE_EXHAUSTED. Raise it well past anything the service itself produces so
# the UDS leg is never the narrower of the two.
_MAX_RECEIVE_MESSAGE_LENGTH = 256 * 1024 * 1024

_CHANNEL_OPTIONS = (("grpc.max_receive_message_length", _MAX_RECEIVE_MESSAGE_LENGTH),)

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Two concerns with the one-sided cap:

1. Send-side left at the gRPC 4 MiB default — silent regression vs. the native path. The native google-cloud-bigtable Python client sets both grpc.max_send_message_length and grpc.max_receive_message_length to -1 (unlimited). This PR raises only inbound, so an ApplyBulk / MutateRow with a payload >4 MiB (allowed by the server's 100 MiB per-cell limit) will fail at the accelerator channel's outbound serialization check with no analogous server-side error. On the native path the same call goes through. Suggest raising send too:

Suggested change
_CHANNEL_OPTIONS = (("grpc.max_receive_message_length", _MAX_RECEIVE_MESSAGE_LENGTH),)
_CHANNEL_OPTIONS = (
("grpc.max_receive_message_length", _MAX_RECEIVE_MESSAGE_LENGTH),
("grpc.max_send_message_length", _MAX_RECEIVE_MESSAGE_LENGTH),
)

2. Pick a side on the 256 MB number. Native uses -1 (unlimited). 256 MB from thin air invites "why not 128? why not 1 GiB?" and future drift when the native client bumps. Either match native (-1, both directions) and document "match native" in the comment, or if 256 MB is a deliberate memory-safety choice at the UDS hop, say that explicitly.



@CrossSync.convert_class(sync_name="_AcceleratorClient")
class _AsyncAcceleratorClient:
Expand All @@ -59,7 +69,9 @@ class _AsyncAcceleratorClient:
def __init__(self, uds_path: str, auth_secret: str):
self._uds_path = uds_path
self._metadata = (("x-accelerator-token", auth_secret),)
self._channel = insecure_channel(_uds_target(uds_path))
self._channel = insecure_channel(
_uds_target(uds_path), options=_CHANNEL_OPTIONS

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nit: no test exercises _CHANNEL_OPTIONS. Even a one-liner asserting the option tuple is on self._channel's options would pin both the key spelling (grpc.max_receive_message_length is easy to typo) and the value. Low priority.

)
self._mutate_row_stub = self._channel.unary_unary(
_MUTATE_ROW_METHOD,
request_serializer=MutateRowRequest.serialize,
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,8 @@

_MUTATE_ROW_METHOD = "/google.bigtable.v2.Bigtable/MutateRow"
_READ_ROWS_METHOD = "/google.bigtable.v2.Bigtable/ReadRows"
_MAX_RECEIVE_MESSAGE_LENGTH = 256 * 1024 * 1024
_CHANNEL_OPTIONS = (("grpc.max_receive_message_length", _MAX_RECEIVE_MESSAGE_LENGTH),)


class _AcceleratorClient:
Expand All @@ -55,7 +57,9 @@ class _AcceleratorClient:
def __init__(self, uds_path: str, auth_secret: str):
self._uds_path = uds_path
self._metadata = (("x-accelerator-token", auth_secret),)
self._channel = insecure_channel(_uds_target(uds_path))
self._channel = insecure_channel(
_uds_target(uds_path), options=_CHANNEL_OPTIONS
)
self._mutate_row_stub = self._channel.unary_unary(
_MUTATE_ROW_METHOD,
request_serializer=MutateRowRequest.serialize,
Expand Down
Loading