Do not marshal the same object twice in `PutSingle` #531

New issue

Closed

opened 2023-07-19 14:48:17 +00:00 by fyrchik · 2 comments

fyrchik commented

2023-07-19 14:48:17 +00:00

Owner

It rubs me wrong way, when I see the same object being repeatedly marshaled just to be sent to different clients. It is the same. Consider PutSingle: we receive a message which contains perfectly valid object, we need to unmarshal it (for validation), but that's all: the same bytes should be both put in the local storage and sent to other clients.
Another thing is payload: marshaled + unmarshaled form take twice as much space as just an object, because the payload is copied (needs verification).

I believe we can do better.
In this task:

Investigate whether there are non-invasive solutions for this kind of task: we do not want to use tricky solutions in the frostfs-api-go.
Investigate whether gRPC codecs could be helpful (raw API client can be used in this single case, because SDK client has typed parameters).

Ideally we would like to have a single slice with the same lifetime as PutSingle RPC (currently its 3x overhead for REP 2, even more for a simple Put).

We will decide what the next step is after research.

Related #463

It rubs me wrong way, when I see the same object being repeatedly marshaled just to be sent to different clients. It _is_ the same. Consider `PutSingle`: we receive a message which contains perfectly valid object, we need to unmarshal it (for validation), but that's all: the same bytes should be both put in the local storage and sent to other clients. Another thing is payload: marshaled + unmarshaled form take twice as much space as just an object, because the payload is copied (needs verification). I believe we can do better. In this task: 1. Investigate whether there are non-invasive solutions for this kind of task: we do not want to use tricky solutions in the frostfs-api-go. 2. Investigate whether gRPC codecs could be helpful (raw API client can be used in this single case, because SDK client has typed parameters). Ideally we would like to have a single slice with the same lifetime as `PutSingle` RPC (currently its 3x overhead for `REP 2`, even more for a simple `Put`). We will decide what the next step is after research. Related #463