[ludfjig]

Rendered

[Copilot]

🟡 Changes recommended

The proposal leaves handling of events larger than 4 KiB undefined and omits the snapshot ABI bump and regeneration requirements needed to prevent incompatible snapshots from being resumed.

2 open issues

• Require snapshot ABI bump for incompatible format changes · New
• Define oversized event handling semantics · New

What changed in this PR

HIP 0008 proposes replacing FlatBuffers with serde and postcard for Hyperlight’s host-guest control messages. This PR contains the design proposal only, not an implementation.

Changes:

• Defines message types, a control-data length field, and trace batching while preserving external byte payloads.
• Describes removing FlatBuffers tooling and generated code.
• Presents preliminary performance results, drawbacks, and alternatives.

File	Description
proposals/​0008-postcard-serialization/​README.md	Documents the serialization proposal, wire format, benchmarks, and tradeoffs.

_{Review effort: Balanced}

[andreiltd]

That was an easy read. Postcard seems like a good candidate for the wire format. IIUC the proposal keeps byte payloads outside the control plane using external bytes, so serialization/deserialization performance isn't much of a concern, as all control messages are expected to be relatively small.

I wonder if you considered the timing for landing this work and fixing: #1987 ?

[ludfjig]

I wonder if you considered the timing for landing this work and fixing: #1987 ?

I think it makes sense to do the serde change first and then #1987 on top, so #1987 doesn't have to change FlatBuffers code that this removes. #1987 then removes the temporary chunked flag along with VecBytes.

[jsturtevant]

who controls this length?

[jsturtevant]

is this an improvement or regression?

[jsturtevant]

couple questions but I like the motivation section, helps understand why we would want to do this!