feat(grpc-events): handle-capture cycle — event-row gap proven NOT a client payload issue

Extends the instrument-grpc rewrite to log string (strHandle) + uint (uiHandle /
queryRequestType) params, not just byte[], and captures our SDK's live v8
openParameters for a byte-diff against the native.

Result of the exhaustive comparison (all live-confirmed via the opt-in
EventReadDiagnostic test):
- StartEventQuery request: byte-identical to the native (v6 layout)
- v8 OpenConnection openParameters: byte-identical to the native (302B) once
  ClientNodeName matches — every control byte/ConnectionType/token/ShardId
- handle usage identical: ExchangeKey->contextKey, registration->storage GUID
  (strHandle), query->client uint (uiHandle); handles valid (RTag/EnsT=True)
- queryRequestType=3, registration order, gzip metadata header — all match
- window has events (native returns 50 now); eventCount not it

Every observable client-side byte matches the native, yet the server scopes 0
events to our connection. The event RPCs succeed over our transport and return a
valid EMPTY result (not a transport error), so this is a connection/server-level
difference (session affinity tied to the native Grpc.Core HTTP/2 connection or a
connection identity used to scope events) — invisible to and unfixable by client
payload matching. Needs server-side insight, not more wire RE.

Added opt-in diagnostics (RegistrationDiag, LastResultBufferHex,
LastEventOpenRequestHex). 326/326 offline; gated test still pins the no-row throw.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B6mcaT2PjRFKcogzp9UkfC
This commit is contained in:
Joseph Doherty
2026-06-23 12:31:33 -04:00
parent 32ae301050
commit 0921e21bdb
5 changed files with 74 additions and 3 deletions
@@ -1387,10 +1387,14 @@ static int InstrumentGrpcNonStream(string[] args)
ModuleDefMD module = ModuleDefMD.Load(sourcePath);
MemberRefUser logByteArray = CreateLogByteArrayRef(module);
MemberRefUser logString = CreateLogStringRef(module);
MemberRefUser logUInt32 = CreateLogUInt32Ref(module);
// Cast a wide net: instrument EVERY byte[]-input method on every Grpc*Client type, so whichever
// path the native non-streamed write actually drives (History/Transaction RegisterTags +
// AddNonStreamValues, or a Storage-service route) is captured. Phase = "<Type>.<Method>.<param>".
// Also captures string (strHandle) and uint (uiHandle / queryRequestType) inputs so the event-read
// handle fields are visible, not just byte[] params.
var instrumented = new List<object>();
foreach (TypeDef type in module.GetTypes()
.Where(t => t.Name.String.StartsWith("Grpc", StringComparison.Ordinal) && t.Name.String.EndsWith("Client", StringComparison.Ordinal)))
@@ -1431,6 +1435,34 @@ static int InstrumentGrpcNonStream(string[] args)
});
}
// ENTRY: log string (strHandle) and uint (uiHandle / queryRequestType / count) inputs.
foreach (dnlib.DotNet.Parameter scalarParam in method.Parameters
.Where(p => !p.IsHiddenThisParameter && (p.Type.FullName == "System.String" || p.Type.FullName == "System.UInt32"))
.ToArray())
{
bool isString = scalarParam.Type.FullName == "System.String";
string phase = $"{type.Name}.{method.Name}.{scalarParam.Name}.{(isString ? "str" : "u32")}";
Instruction[] injected =
[
Instruction.Create(OpCodes.Ldstr, phase),
Instruction.Create(OpCodes.Ldarg, scalarParam),
Instruction.Create(OpCodes.Call, isString ? logString : logUInt32),
];
foreach (Instruction instruction in injected.Reverse())
{
method.Body.Instructions.Insert(0, instruction);
}
method.Body.MaxStack = (ushort)Math.Max((int)method.Body.MaxStack, 8);
instrumented.Add(new
{
Type = type.Name.String,
Method = method.Name.String,
Phase = phase,
Direction = "in",
Token = "0x" + method.MDToken.Raw.ToString("X8"),
});
}
// EXIT: log out/ref byte[] responses ("System.Byte[]&") before each ret. ldarg loads the
// managed pointer; ldind.ref dereferences it to the byte[]. (RPC wrappers set the out
// param right before a single ret, so branch-to-ret skew is not a concern here.)