Enable using gRPC client for GCS #92
Conversation
GCP has an experimental feature which allows using gRPC when talking to GCS. For the time being, it requires buckets to be explicitly added to an allowlist, but eventually the feature will become GA. This commit allows using a gRPC client when creating a GCS bucket by toggling a boolean flag in the bucket config. Signed-off-by: Filip Petkovski <[email protected]>
Signed-off-by: Filip Petkovski <[email protected]>
fpetkovski
force-pushed
the
gcs-grpc-client
branch
from
3e72eef to
c035baa
Compare
Signed-off-by: Filip Petkovski <[email protected]>
providers/gcs/gcs.go
Outdated
| Bucket string `yaml:"bucket"` | ||
| ServiceAccount string `yaml:"service_account"` | ||
| UseGRPC bool `yaml:"use_grpc"` | ||
| GRPCConnPoolSize int `yaml:"grpc_conn_pool_size"` |
There was a problem hiding this comment.
I did a bit of testing by tweaking this parameter up and down. Performance follows a U-shaped curve where using 1 connection is abysmally slow, even slower than HTTP. With a pool size of 512 performance is better, but it seems to be best with 256.
We could consider using that (or some other number) as the default to make sure users do not need to guess in the beginning. We can still keep the value configurable if anyone needs further customization.
There was a problem hiding this comment.
Turns out we had to enable direct path as documented here https://cloud.google.com/go/docs/reference/cloud.google.com/go/storage/latest#hdr-Experimental_gRPC_API.
If the application is running within GCP, users may get better performance by enabling DirectPath (enabling requests to skip some proxy steps). To enable, set the environment variable GOOGLE_CLOUD_ENABLE_DIRECT_PATH_XDS=true and add the following side-effect imports to your application:
import (
_ "google.golang.org/grpc/balancer/rls"
_ "google.golang.org/grpc/xds/googledirectpath"
)This made the connection pool config obsolete (as a matter of fact, anything higher than 1 would hang the client). The config option is still useful when direct path is not enabled though.
There was a problem hiding this comment.
We can probably mention the direct path option as a comment?
There was a problem hiding this comment.
Good point, updated.
|
@saswatamcode any thoughts on this one? |
providers/gcs/gcs.go
Outdated
| Bucket string `yaml:"bucket"` | ||
| ServiceAccount string `yaml:"service_account"` | ||
| UseGRPC bool `yaml:"use_grpc"` | ||
| GRPCConnPoolSize int `yaml:"grpc_conn_pool_size"` |
There was a problem hiding this comment.
We can probably mention the direct path option as a comment?
Signed-off-by: Filip Petkovski <[email protected]>
GCP has an experimental feature which allows using gRPC when talking to GCS. For the time being, it requires buckets to be explicitly added to an allowlist, but eventually the feature will become GA.
This commit allows using a gRPC client when creating a GCS bucket by toggling a boolean flag in the bucket config.
Changes
Verification