reports.proto

syntax = "proto3";

// package mdg.engine.proto;

import "google/protobuf/timestamp.proto";

message Trace {
	message CachePolicy {
		enum Scope {
			UNKNOWN = 0;
			PUBLIC = 1;
			PRIVATE = 2;
		}

		Scope scope = 1;
		int64 max_age_ns = 2; // use 0 for absent, -1 for 0
	}

	message Details {
		// The variables associated with this query (unless the reporting agent is
		// configured to keep them all private). Values are JSON: ie, strings are
		// enclosed in double quotes, etc.  The value of a private variable is
		// the empty string.
		map<string, string> variables_json = 4;
		// Deprecated. Engineproxy did not encode variable values as JSON, so you
		// couldn't tell numbers from numeric strings. Send variables_json instead.
		map<string, bytes> variables = 1;
		// Optional: this is the original full query before the signature algorithm
		// is applied.  Engineproxy always sent this in all traces, which meant that
		// literal-masking done by the signature algorithm didn't fully hide
		// sensitive data from Engine servers. apollo-engine-reporting does not
		// include this by default.  (The Engine frontend does not currently show
		// this field.)
		string raw_query = 2;
		// Don't include this in traces inside a FullTracesReport; the operation
		// name for these traces comes from the key of the traces_per_query map.
		string operation_name = 3;
	}

	message Error {
		string message = 1; // required
		repeated Location location = 2;
		uint64 time_ns = 3;
		string json = 4;
	}

	message HTTP {
		message Values {
			repeated string value = 1;
		}

		enum Method {
			UNKNOWN = 0;
			OPTIONS = 1;
			GET = 2;
			HEAD = 3;
			POST = 4;
			PUT = 5;
			DELETE = 6;
			TRACE = 7;
			CONNECT = 8;
			PATCH = 9;
		}
		Method method = 1;
		string host = 2;
		string path = 3;

		// Should exclude manual blacklist ("Auth" by default)
		map<string, Values> request_headers = 4;
		map<string, Values> response_headers = 5;

		uint32 status_code = 6;

		bool secure = 8; // TLS was used
		string protocol = 9; // by convention "HTTP/1.0", "HTTP/1.1", "HTTP/2" or "h2"
	}

	message Location {
		uint32 line = 1;
		uint32 column = 2;
	}

	// We store information on each resolver execution as a Node on a tree.
	// The structure of the tree corresponds to the structure of the GraphQL
	// response; it does not indicate the order in which resolvers were
	// invoked.  Note that nodes representing indexes (and the root node)
	// don't contain all Node fields (eg types and times).
	message Node {
		// The name of the field (for Nodes representing a resolver call) or the
		// index in a list (for intermediate Nodes representing elements of a list).
		// field_name is the name of the field as it appears in the GraphQL
		// response: ie, it may be an alias.  (In that case, the original_field_name
		// field holds the actual field name from the schema.) In any context where
		// we're building up a path, we use the response_name rather than the
		// original_field_name.
		oneof id {
			string response_name = 1;
			uint32 index = 2;
		}

		string original_field_name = 14;

		// The field's return type; e.g. "String!" for User.email:String!
		string type = 3;

		// The field's parent type; e.g. "User" for User.email:String!
		string parent_type = 13;

		CachePolicy cache_policy = 5;

		// relative to the trace's start_time, in ns
		uint64 start_time = 8;
		// relative to the trace's start_time, in ns
		uint64 end_time = 9;

		repeated Error error = 11;
		repeated Node child = 12;

		reserved 4;
	}

	// represents a node in the query plan, under which there is a trace tree for that service fetch.
	// In particular, each fetch node represents a call to an implementing service, and calls to implementing
	// services may not be unique. See https://github.com/apollographql/apollo-server/blob/master/packages/apollo-gateway/src/QueryPlan.ts
	// for more information and details.
	message QueryPlanNode {
		// This represents a set of nodes to be executed sequentially by the Gateway executor
		message SequenceNode {
			repeated QueryPlanNode nodes = 1;
		}
		// This represents a set of nodes to be executed in parallel by the Gateway executor
		message ParallelNode {
			repeated QueryPlanNode nodes = 1;
		}
		// This represents a node to send an operation to an implementing service
		message FetchNode {
			// XXX When we want to include more details about the sub-operation that was
			// executed against this service, we should include that here in each fetch node.
			// This might include an operation signature, requires directive, reference resolutions, etc.
			string serviceName = 1;

			bool traceParsingFailed = 2;

			// This Trace only contains start_time, end_time, duration_ns, and root;
			// all timings were calculated **on the federated service**, and clock skew
			// will be handled by the ingress server.
			Trace trace = 3;

			// relative to the outer trace's start_time, in ns, measured in the gateway.
			uint64 sent_time_offset = 4;

			// Wallclock times measured in the gateway for when this operation was
			// sent and received.
			google.protobuf.Timestamp sent_time = 5;
			google.protobuf.Timestamp received_time = 6;
		}

		// This node represents a way to reach into the response path and attach related entities.
		// XXX Flatten is really not the right name and this node may be renamed in the query planner.
		message FlattenNode {
			repeated ResponsePathElement response_path = 1;
			QueryPlanNode node = 2;
		}
		message ResponsePathElement {
			oneof id {
				string field_name = 1;
				uint32 index = 2;
			}
		}
		oneof node {
			SequenceNode sequence = 1;
			ParallelNode parallel = 2;
			FetchNode fetch = 3;
			FlattenNode flatten = 4;
		}
	}

	// Wallclock time when the trace began.
	google.protobuf.Timestamp start_time = 4; // required
	// Wallclock time when the trace ended.
	google.protobuf.Timestamp end_time = 3; // required
	// High precision duration of the trace; may not equal end_time-start_time
	// (eg, if your machine's clock changed during the trace).
	uint64 duration_ns = 11; // required
	// A tree containing information about all resolvers run directly by this
	// service, including errors.
	Node root = 14;

	// -------------------------------------------------------------------------
	// Fields below this line are *not* included in federated traces (the traces
	// sent from federated services to the gateway).

	// In addition to details.raw_query, we include a "signature" of the query,
	// which can be normalized: for example, you may want to discard aliases, drop
	// unused operations and fragments, sort fields, etc. The most important thing
	// here is that the signature match the signature in StatsReports. In
	// StatsReports signatures show up as the key in the per_query map (with the
	// operation name prepended).  The signature should be a valid GraphQL query.
	// All traces must have a signature; if this Trace is in a FullTracesReport
	// that signature is in the key of traces_per_query rather than in this field.
	// Engineproxy provides the signature in legacy_signature_needs_resigning
	// instead.
	string signature = 19;

	Details details = 6;

	// Note: engineproxy always sets client_name, client_version, and client_address to "none".
	// apollo-engine-reporting allows for them to be set by the user.
	string client_name = 7;
	string client_version = 8;
	string client_address = 9;
	string client_reference_id = 23;

	HTTP http = 10;

	CachePolicy cache_policy = 18;

	// If this Trace was created by a gateway, this is the query plan, including
	// sub-Traces for federated services. Note that the 'root' tree on the
	// top-level Trace won't contain any resolvers (though it could contain errors
	// that occurred in the gateway itself).
	QueryPlanNode query_plan = 26;

	// Was this response served from a full query response cache?  (In that case
	// the node tree will have no resolvers.)
	bool full_query_cache_hit = 20;

	// Was this query specified successfully as a persisted query hash?
	bool persisted_query_hit = 21;
	// Did this query contain both a full query string and a persisted query hash?
	// (This typically means that a previous request was rejected as an unknown
	// persisted query.)
	bool persisted_query_register = 22;

	// Was this operation registered and a part of the safelist?
	bool registered_operation = 24;

	// Was this operation forbidden due to lack of safelisting?
	bool forbidden_operation = 25;

	// --------------------------------------------------------------
	// Fields below this line are only set by the old Go engineproxy.
	google.protobuf.Timestamp origin_reported_start_time = 15;
	google.protobuf.Timestamp origin_reported_end_time = 16;
	uint64 origin_reported_duration_ns = 17;

	// Older agents (eg the Go engineproxy) relied to some degree on the Engine
	// backend to run their own semi-compatible implementation of a specific
	// variant of query signatures. The backend does not do this for new agents (which
	// set the above 'signature' field). It used to still "re-sign" signatures
	// from engineproxy, but we've now simplified the backend to no longer do this.
	// Deprecated and ignored in FullTracesReports.
	string legacy_signature_needs_resigning = 5;


	// removed: Node parse = 12; Node validate = 13;
	//          Id128 server_id = 1; Id128 client_id = 2;
	reserved 12, 13, 1, 2;
}

// The `service` value embedded within the header key is not guaranteed to contain an actual service,
// and, in most cases, the service information is trusted to come from upstream processing. If the
// service _is_ specified in this header, then it is checked to match the context that is reporting it.
// Otherwise, the service information is deduced from the token context of the reporter and then sent
// along via other mechanisms (in Kafka, the `ReportKafkaKey). The other information (hostname,
// agent_version, etc.) is sent by the Apollo Engine Reporting agent, but we do not currently save that
// information to any of our persistent storage.
message ReportHeader {
	string service = 3;
	// eg "host-01.example.com"
	string hostname = 5;

	// eg "engineproxy 0.1.0"
	string agent_version = 6; // required
	// eg "prod-4279-20160804T065423Z-5-g3cf0aa8" (taken from `git describe --tags`)
	string service_version = 7;
	// eg "node v4.6.0"
	string runtime_version = 8;
	// eg "Linux box 4.6.5-1-ec2 #1 SMP Mon Aug 1 02:31:38 PDT 2016 x86_64 GNU/Linux"
	string uname = 9;
	// eg "current", "prod"
	string schema_tag = 10;
	// The hex representation of the sha512 of the introspection response
	string schema_hash = 11;
}

message PathErrorStats {
	map<string, PathErrorStats> children = 1;
	uint64 errors_count = 4;
	uint64 requests_with_errors_count = 5;
}

message ClientNameStats {
	// Duration histogram for non-cache-hit queries.
	// (See docs/histograms.md for the histogram format.)
	repeated int64 latency_count = 1;
	reserved 2; // removed: repeated uint64 error_count = 2;
	// These per-version fields were used to understand what versions contributed to this sample
	// when we were implementing the aggregation of this information ourselves using BigTable.
	// However, since the per-version stats don't separate out latency, it makes more sense to
	// have stats reported with contextual information so we can have the specific breakdown we're
	// looking for. These fields are somewhat misleading as we never actually do any per-version
	// awareness with anything reporting in the legacy "per_client_name" stats, and instead use
	// "query_stats_with_context" to have more contextual information.
	map<string, uint64> requests_count_per_version = 3; // required
	map<string, uint64> cache_hits_per_version = 4;
	map<string, uint64> persisted_query_hits_per_version = 10;
	map<string, uint64> persisted_query_misses_per_version = 11;
	map<string, uint64> registered_operation_count_per_version = 12;
	map<string, uint64> forbidden_operation_count_per_version = 13;
	repeated int64 cache_latency_count = 5; // Duration histogram; see docs/histograms.md
	PathErrorStats root_error_stats = 6;
	uint64 requests_with_errors_count = 7;
	// TTL histograms for cache misses for the public cache.
	repeated int64 public_cache_ttl_count = 8;
	// TTL histograms for cache misses for the private cache.
	repeated int64 private_cache_ttl_count = 9;
}

message QueryLatencyStats {
	repeated int64 latency_count = 1;
	uint64 request_count = 2;
	uint64 cache_hits = 3;
	uint64 persisted_query_hits = 4;
	uint64 persisted_query_misses = 5;
	repeated int64 cache_latency_count = 6;
	PathErrorStats root_error_stats = 7;
	uint64 requests_with_errors_count = 8;
	repeated int64 public_cache_ttl_count = 9;
	repeated int64 private_cache_ttl_count = 10;
	uint64 registered_operation_count = 11;
	uint64 forbidden_operation_count = 12;
}

message StatsContext {
	string client_reference_id = 1;
	string client_name = 2;
	string client_version = 3;
}

message ContextualizedQueryLatencyStats {
	QueryLatencyStats query_latency_stats = 1;
	StatsContext context = 2;
}

message ContextualizedTypeStats {
	StatsContext context = 1;
	map<string, TypeStat> per_type_stat = 2;
}

message FieldStat {
	string name = 2; // deprecated; only set when stored in TypeStat.field
	string return_type = 3; // required; eg "String!" for User.email:String!
	uint64 errors_count = 4;
	uint64 count = 5;
	uint64 requests_with_errors_count = 6;
	repeated int64 latency_count = 8; // Duration histogram; see docs/histograms.md
}

message TypeStat {
	string name = 1; // deprecated; only set when stored in QueryStats.per_type
	repeated FieldStat field = 2;  // deprecated; use per_field_stat instead
	// Key is (eg) "email" for User.email:String!
	map<string, FieldStat> per_field_stat = 3;
}

message QueryStats {
	// Either per_client_name (for back-compat) or query_stats_with_context must be specified. If both are
	// specified, then query_stats_with_context will be used and per_client_name will be ignored. Although
	// the fields in ClientNameStats mention things "per-version," the information in the "per-version"
	// fields will only ever be over the default version, the empty String: "", if arrived at via the
	// FullTracesAggregator.
	map<string, ClientNameStats> per_client_name = 1; // deprecated; use stats_with_context instead
	repeated ContextualizedQueryLatencyStats query_stats_with_context = 4;
	repeated TypeStat per_type = 2; // deprecated; use type_stats_with_context instead
	// Key is the parent type, e.g. "User" for User.email:String!
	map<string, TypeStat> per_type_stat = 3; // deprecated; use type_stats_with_context instead
	repeated ContextualizedTypeStats type_stats_with_context = 5;
}

// Top-level message type for the server-side traces endpoint
message TracesReport {
	ReportHeader header = 1; // required
	repeated Trace trace = 2; // required
}

message Field {
	string name = 2; // required; eg "email" for User.email:String!
	string return_type = 3; // required; eg "String!" for User.email:String!
}

message Type {
	string name = 1; // required; eg "User" for User.email:String!
	repeated Field field = 2;
}

message MemStats {
	uint64 total_bytes = 1; // MemStats.Sys
	uint64 stack_bytes = 2; // MemStats.StackSys
	uint64 heap_bytes = 3; // MemStats.HeapSys
	uint64 heap_released_bytes = 13; // MemStats.HeapReleased
	uint64 gc_overhead_bytes = 4; // MemStats.GCSys

	uint64 stack_used_bytes = 5; // MemStats.StackInuse
	uint64 heap_allocated_bytes = 6; // MemStats.HeapAlloc
	uint64 heap_allocated_objects = 7; // MemStats.HeapObjects

	uint64 heap_allocated_bytes_delta = 8; // MemStats.TotalAlloc delta
	uint64 heap_allocated_objects_delta = 9; // MemStats.Mallocs delta
	uint64 heap_freed_objects_delta = 10; // MemStats.Frees delta

	uint64 gc_stw_ns_delta = 11; // MemStats.PauseTotalNs delta
	uint64 gc_count_delta = 12; // MemStats.NumGC delta
}

message TimeStats {
	uint64 uptime_ns = 1;
	uint64 real_ns_delta = 2;
	uint64 user_ns_delta = 3;
	uint64 sys_ns_delta = 4;
}

// Top-level message type for the server-side stats endpoint
message StatsReport {
	ReportHeader header = 1; // required

	// These fields are about properties of the engineproxy and are not generated
	// from FullTracesReports.
	MemStats mem_stats = 2;
	TimeStats time_stats = 3;

	// Beginning of the period over which stats are collected.
	google.protobuf.Timestamp start_time = 8;
	// End of the period of which stats are collected.
	google.protobuf.Timestamp end_time = 9;
	// Only used to interpret mem_stats and time_stats; not generated from
	// FullTracesReports.
	uint64 realtime_duration = 10;


	// Maps from query descriptor to QueryStats. Required unless
	// legacy_per_query_missing_operation_name is set. The keys are strings of the
	// form `# operationName\nsignature` (literal hash and space), with
	// operationName - if there is no operation name.
	map<string, QueryStats> per_query = 14;

	// Older agents (Go engineproxy) didn't explicitly include the operation name
	// in the key of this map, and the server had to parse it out (after a
	// re-signing operation which is no longer performed). The key here is just the query
	// signature. Deprecated.
	map<string, QueryStats> legacy_per_query_implicit_operation_name = 12;

	// Deprecated: it was useful in Optics where we had access to the whole schema
	// but has not been ever used in Engine.  apollo-engine-reporting will not
	// send it.
	repeated Type type = 13;
}

// This is the top-level message used by the new traces ingress. This
// is designed for the apollo-engine-reporting TypeScript agent and will
// eventually be documented as a public ingress API. This message consists
// solely of traces; the equivalent of the StatsReport is automatically
// generated server-side from this message.  Agents should send traces
// for all requests in this report.  Generally, buffering up until a large
// size has been reached (say, 4MB) or 5-10 seconds has passed is appropriate.
message FullTracesReport {
	ReportHeader header = 1;

	// key is statsReportKey (# operationName\nsignature) Note that the nested
	// traces will *not* have a signature or details.operationName (because the
	// key is adequate).
	//
	// We also assume that traces don't have
	// legacy_per_query_implicit_operation_name, and we don't require them to have
	// details.raw_query (which would consume a lot of space and has privacy/data
	// access issues, and isn't currently exposed by our app anyway).
	map<string, Traces> traces_per_query = 5;
}

// Just a sequence of traces with the same statsReportKey.
message Traces {
	repeated Trace trace = 1;
}

message TraceV1 {
	ReportHeader header = 1;
	Trace trace = 2;
}