KU Leuven

robin.marx@kuleuven.be

Facebook

lniccolini@fb.com

Protocol Labs

marten@protocol.ai

Transport QUIC Internet-Draft This document describes a high-level schema for a standardized logging format called qlog. This format allows easy sharing of data and the creation of reusable visualization and debugging tools. The high-level schema in this document is intended to be protocol-agnostic. Separate documents specify how the format should be used for specific protocol data. The schema is also format-agnostic, and can be represented in for example JSON, csv or protobuf.

There is currently a lack of an easily usable, standardized endpoint logging format. Especially for the use case of debugging and evaluating modern Web protocols and their performance, it is often difficult to obtain structured logs that provide adequate information for tasks like problem root cause analysis. This document aims to provide a high-level schema and harness that describes the general layout of an easily usable, shareable, aggregatable and structured logging format. This high-level schema is protocol agnostic, with logging entries for specific protocols and use cases being defined in other documents (see for example for QUIC and for HTTP/3 and QPACK-related event definitions). The goal of this high-level schema is to provide amenities and default characteristics that each logging file should contain (or should be able to contain), such that generic and reusable toolsets can be created that can deal with logs from a variety of different protocols and use cases. As such, this document contains concepts such as versioning, metadata inclusion, log aggregation, event grouping and log file size reduction techniques. Feedback and discussion are welcome at https://github.com/quiclog/internet-drafts. Readers are advised to refer to the “editor’s draft” at that URL for an up-to-date version of this document.

The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in . While the qlog schema’s are format-agnostic, for readability the qlog documents will use a JSON-inspired format () for examples and definitions. As qlog can be serialized both textually but also in binary, we employ a custom datatype definition language, inspired loosely by the “TypeScript” language. This document describes how to employ JSON and NDJSON as textual serializations for qlog in . Other documents will describe how to utilize other concrete serialization options, though tips and requirements for these are also listed in this document (). The main general conventions in this document a reader should be aware of are: obj? : this object is optional type1 | type2 : a union of these two types (object can be either type1 OR type2) obj:type : this object has this concrete type obj:array<type> : this object is an array of this type class : defines a new type // : single-line comment The main data types are: int8 : signed 8-bit integer int16 : signed 16-bit integer int32 : signed 32-bit integer int64 : signed 64-bit integer uint8 : unsigned 8-bit integer uint16 : unsigned 16-bit integer uint32 : unsigned 32-bit integer uint64 : unsigned 64-bit integer float : 32-bit floating point value double : 64-bit floating point value byte : an individual raw byte (8-bit) value (use array<byte> or the shorthand “bytes” to specify a binary blob) string : list of Unicode (typically UTF-8) encoded characters boolean : boolean enum: fixed list of values (Unless explicity defined, the value of an enum entry is the string version of its name (e.g., initial = “initial”)) any : represents any object type. Mainly used here as a placeholder for more concrete types defined in related documents (e.g., specific event types) All timestamps and time-related values (e.g., offsets) in qlog are logged as doubles in the millisecond resolution. Other qlog documents can define their own data types (e.g., separately for each Packet type that a protocol supports).

The main tenets for the qlog schema design are: Streamable, event-based logging Flexibility in the format, complexity in the tooling (e.g., few components are a MUST, tools need to deal with this) Extensible and pragmatic (e.g., no complex fixed schema with extension points) Aggregation and transformation friendly (e.g., the top-level element is a container for individual traces, group_id can be used to tag events to a particular context) Metadata is stored together with event data

A qlog file should be able to contain several indivdual traces and logs from multiple vantage points that are in some way related. To that end, the top-level element in the qlog schema defines only a small set of “header” fields and an array of component traces. For this document, the required “qlog_version” field MUST have a value of “qlog-03-WIP”. there have been several previously broadly deployed qlog versions based on older drafts of this document (see draft-marx-qlog-main-schema). The old values for the “qlog_version” field were “draft-00”, “draft-01” and “draft-02”. When qlog was moved to the QUIC working group, we decided to increment the existing counter, rather than reverting back to -00. As such, any numbering indicating in the “qlog_version” field is explicitly not tied to a particular version of the draft documents. As qlog can be serialized in a variety of ways, the “qlog_format” field is used to indicate which serialization option was chosen. Its value MUST either be one of the options defined in this document (e.g., ) or the field must be omitted entirely, in which case it assumes the default value of “JSON”. In order to make it easier to parse and identify qlog files and their serialization format, the “qlog_version” and “qlog_format” fields and their values SHOULD be in the first 256 characters/bytes of the resulting log file. An example of the qlog file’s top-level structure is shown in .

} JSON serialization: { "qlog_version": "draft-03-WIP", "qlog_format": "JSON", "title": "Name of this particular qlog file (short)", "description": "Description for this group of traces (long)", "summary": { ... }, "traces": [...] } ]]>

In a real-life deployment with a large amount of generated logs, it can be useful to sort and filter logs based on some basic summarized or aggregated data (e.g., log length, packet loss rate, log location, presence of error events, …). The summary field (if present) SHOULD be on top of the qlog file, as this allows for the file to be processed in a streaming fashion (i.e., the implementation could just read up to and including the summary field and then only load the full logs that are deemed interesting by the user). As the summary field is highly deployment-specific, this document does not specify any default fields or their semantics. Some examples of potential entries are shown in .

It is often advantageous to group several related qlog traces together in a single file. For example, we can simultaneously perform logging on the client, on the server and on a single point on their common network path. For analysis, it is useful to aggregate these three individual traces together into a single file, so it can be uniquely stored, transferred and annotated. As such, the “traces” array contains a list of individual qlog traces. Typical qlogs will only contain a single trace in this array. These can later be combined into a single qlog file by taking the “traces” entry/entries for each qlog file individually and copying them to the “traces” array of a new, aggregated qlog file. This is typically done in a post-processing step. The “traces” array can thus contain both normal traces (for the definition of the Trace type, see ), but also “error” entries. These indicate that we tried to find/convert a file for inclusion in the aggregated qlog, but there was an error during the process. Rather than silently dropping the erroneous file, we can opt to explicitly include it in the qlog file as an entry in the “traces” array, as shown in .

Note that another way to combine events of different traces in a single qlog file is through the use of the “group_id” field, discussed in .

The exact conceptual definition of a Trace can be fluid. For example, a trace could contain all events for a single connection, for a single endpoint, for a single measurement interval, for a single protocol, etc. As such, a Trace container contains some metadata in addition to the logged events, see . In the normal use case however, a trace is a log of a single data flow collected at a single location or vantage point. For example, for QUIC, a single trace only contains events for a single logical QUIC connection for either the client or the server. The semantics and context of the trace can mainly be deduced from the entries in the “common_fields” list and “vantage_point” field.

} JSON serialization: { "title": "Name of this particular trace (short)", "description": "Description for this trace (long)", "configuration": { "time_offset": 150 }, "common_fields": { "ODCID": "abcde1234", "time_format": "absolute" }, "vantage_point": { "name": "backend-67", "type": "server" }, "events": [...] } ]]>

We take into account that a qlog file is usually not used in isolation, but by means of various tools. Especially when aggregating various traces together or preparing traces for a demonstration, one might wish to persist certain tool-based settings inside the qlog file itself. For this, the configuration field is used. The configuration field can be viewed as a generic metadata field that tools can fill with their own fields, based on per-tool logic. It is best practice for tools to prefix each added field with their tool name to prevent collisions across tools. This document only defines two optional, standard, tool-independent configuration settings: “time_offset” and “original_uris”.

, // list of fields with any type } JSON serialization: { "time_offset": 150, // starts 150ms after the first timestamp indicates "original_uris": [ "https://example.org/trace1.qlog", "https://example.org/trace2.qlog" ] } ]]>

The time_offset field indicates by how many milliseconds the starting time of the current trace should be offset. This is useful when comparing logs taken from various systems, where clocks might not be perfectly synchronous. Users could use manual tools or automated logic to align traces in time and the found optimal offsets can be stored in this field for future usage. The default value is 0.

The original_uris field is used when merging multiple individual qlog files or other source files (e.g., when converting .pcaps to qlog). It allows to keep better track where certain data came from. It is a simple array of strings. It is an array instead of a single string, since a single qlog trace can be made up out of an aggregation of multiple component qlog traces as well. The default value is an empty array.

Tools can add optional custom metadata to the “configuration” field to store state and make it easier to share specific data viewpoints and view configurations. Two examples from the qvis toolset are shown in .

The vantage_point field describes the vantage point from which the trace originates, see . Each trace can have only a single vantage_point and thus all events in a trace MUST BE from the perspective of this vantage_point. To include events from multiple vantage_points, implementers can for example include multiple traces, split by vantage_point, in a single qlog file.

The flow field is only required if the type is “network” (for example, the trace is generated from a packet capture). It is used to disambiguate events like “packet sent” and “packet received”. This is indicated explicitly because for multiple reasons (e.g., privacy) data from which the flow direction can be otherwise inferred (e.g., IP addresses) might not be present in the logs. Meaning of the different values for the flow field: * “client” indicates that this vantage point follows client data flow semantics (a “packet sent” event goes in the direction of the server). * “server” indicates that this vantage point follow server data flow semantics (a “packet sent” event goes in the direction of the client). * “unknown” indicates that the flow’s direction is unknown. Depending on the context, tools confronted with “unknown” values in the vantage_point can either try to heuristically infer the semantics from protocol-level domain knowledge (e.g., in QUIC, the client always sends the first packet) or give the user the option to switch between client and server perspectives manually.

Inside of the “events” field of a qlog trace is a list of events logged by the endpoint. Each event is specified as a generic object with a number of member fields and their associated data. Depending on the protocol and use case, the exact member field names and their formats can differ across implementations. This section lists the main, pre-defined and reserved field names with specific semantics and expected corresponding value formats. Each qlog event at minimum requires the “time” (), “name” () and “data” () fields. Other typical fields are “time_format” (), “protocol_type” (), “trigger” (), and “group_id” . As especially these later fields typically have identical values across individual event instances, they are normally logged separately in the “common_fields” (). The specific values for each of these fields and their semantics are defined in separate documents, specific per protocol or use case. For example: event definitions for QUIC, HTTP/3 and QPACK can be found in and . Other fields are explicitly allowed by the qlog approach, and tools SHOULD allow for the presence of unknown event fields, but their semantics depend on the context of the log usage (e.g., for QUIC, the ODCID field is used), see . An example of a qlog event with its component fields is shown in .

, group_id?: string|uint32, time_format?: "absolute"|"delta"|"relative", // list of fields with any type } JSON serialization: { time: 1553986553572, name: "transport:packet_sent", data: { ... } protocol_type: ["QUIC","HTTP3"], group_id: "127ecc830d98f9d54a42c4f0842aa87e181a", time_format: "absolute", ODCID: "127ecc830d98f9d54a42c4f0842aa87e181a", // QUIC specific } ]]>

The “time” field indicates the timestamp at which the event occured. Its value is typically the Unix timestamp since the 1970 epoch (number of milliseconds since midnight UTC, January 1, 1970, ignoring leap seconds). However, qlog supports two more succint timestamps formats to allow reducing file size. The employed format is indicated in the “time_format” field, which allows one of three values: “absolute”, “delta” or “relative”: Absolute: Include the full absolute timestamp with each event. This approach uses the largest amount of characters. This is also the default value of the “time_format” field. Delta: Delta-encode each time value on the previously logged value. The first event in a trace typically logs the full absolute timestamp. This approach uses the least amount of characters. Relative: Specify a full “reference_time” timestamp (typically this is done up-front in “common_fields”, see ) and include only relatively-encoded values based on this reference_time with each event. The “reference_time” value is typically the first absolute timestamp. This approach uses a medium amount of characters. The first option is good for stateless loggers, the second and third for stateful loggers. The third option is generally preferred, since it produces smaller files while being easier to reason about. An example for each option can be seen in .

One of these options is typically chosen for the entire trace (put differently: each event has the same value for the “time_format” field). Each event MUST include a timestamp in the “time” field. Events in each individual trace SHOULD be logged in strictly ascending timestamp order (though not necessarily absolute value, for the “delta” format). Tools CAN sort all events on the timestamp before processing them, though are not required to (as this could impose a significant processing overhead). This can be a problem especially for multi-threaded and/or streaming loggers, who could consider using a separate postprocesser to order qlog events in time if a tool do not provide this feature. Timestamps do not have to use the UNIX epoch timestamp as their reference. For example for privacy considerations, any initial reference timestamps (for example “endpoint uptime in ms” or “time since connection start in ms”) can be chosen. Tools SHOULD NOT assume the ability to derive the absolute Unix timestamp from qlog traces, nor allow on them to relatively order events across two or more separate traces (in this case, clock drift should also be taken into account).

Events differ mainly in the type of metadata associated with them. To help identify a given event and how to interpret its metadata in the “data” field (see ), each event has an associated “name” field. This can be considered as a concatenation of two other fields, namely event “category” and event “type”. Category allows a higher-level grouping of events per specific event type. For example for QUIC and HTTP/3, the different categories could be “transport”, “http”, “qpack”, and “recovery”. Within these categories, the event Type provides additional granularity. For example for QUIC and HTTP/3, within the “transport” Category, there would be “packet_sent” and “packet_received” events. Logging category and type separately conceptually allows for fast and high-level filtering based on category and the re-use of event types across categories. However, it also considerably inflates the log size and this flexibility is not used extensively in practice at the time of writing. As such, the default approach in qlog is to concatenate both field values using the “:” character in the “name” field, as can be seen in . As such, qlog category and type names MUST NOT include this character.

Certain serializations CAN emit category and type as separate fields, and qlog tools SHOULD be able to deal with both the concatenated “name” field, and the separate “category” and “type” fields. Text-based serializations however are encouraged to employ the concatenated “name” field for efficiency.

The data field is a generic object. It contains the per-event metadata and its form and semantics are defined per specific sort of event. For example, data field value definitons for QUIC and HTTP/3 can be found in and . One purely illustrative example for a QUIC “packet_sent” event is shown in .

} JSON serialization: { packet_size: 1280, header: { packet_type: "1RTT", packet_number: 123 }, frames: [ { frame_type: "stream", length: 1000, offset: 456 }, { frame_type: "padding" } ] } ]]>

The “protocol_type” array field indicates to which protocols (or protocol “stacks”) this event belongs. This allows a single qlog file to aggregate traces of different protocols (e.g., a web server offering both TCP+HTTP/2 and QUIC+HTTP/3 connections). For example, QUIC and HTTP/3 events have the “QUIC” and “HTTP3” protocol_type entry values, see and . Typically however, all events in a single trace are of the same few protocols, and this array field is logged once in “common_fields”, see .

Sometimes, additional information is needed in the case where a single event can be caused by a variety of other events. In the normal case, the context of the surrounding log messages gives a hint as to which of these other events was the cause. However, in highly-parallel and optimized implementations, corresponding log messages might separated in time. Another option is to explicitly indicate these “triggers” in a high-level way per-event to get more fine-grained information without much additional overhead. In qlog, the optional “trigger” field contains a string value describing the reason (if any) for this event instance occuring. While this “trigger” field could be a property of the qlog Event itself, it is instead a property of the “data” field instead. This choice was made because many event types do not include a trigger value, and having the field at the Event-level would cause overhead in some serializations. Additional information on the trigger can be added in the form of additional member fields of the “data” field value, yet this is highly implementation-specific, as are the trigger field’s string values. One purely illustrative example of some potential triggers for QUIC’s “packet_dropped” event is shown in .

As discussed in , a single qlog file can contain several traces taken from different vantage points. However, a single trace from one endpoint can also contain events from a variety of sources. For example, a server implementation might choose to log events for all incoming connections in a single large (streamed) qlog file. As such, we need a method for splitting up events belonging to separate logical entities. The simplest way to perform this splitting is by associating a “group identifier” to each event that indicates to which conceptual “group” each event belongs. A post-processing step can then extract events per group. However, this group identifier can be highly protocol and context-specific. In the example above, we might use QUIC’s “Original Destination Connection ID” to uniquely identify a connection. As such, they might add a “ODCID” field to each event. However, a middlebox logging IP or TCP traffic might rather use four-tuples to identify connections, and add a “four_tuple” field. As such, to provide consistency and ease of tooling in cross-protocol and cross-context setups, qlog instead defines the common “group_id” field, which contains a string value. Implementations are free to use their preferred string serialization for this field, so long as it contains a unique value per logical group. Some examples can be seen in .

Note that in some contexts (for example a Multipath transport protocol) it might make sense to add additional contextual per-event fields (for example “path_id”), rather than use the group_id field for that purpose. Note also that, typically, a single trace only contains events belonging to a single logical group (for example, an individual QUIC connection). As such, instead of logging the “group_id” field with an identical value for each event instance, this field is typically logged once in “common_fields”, see .

As discussed in the previous sections, information for a typical qlog event varies in three main fields: “time”, “name” and associated data. Additionally, there are also several more advanced fields that allow mixing events from different protocols and contexts inside of the same trace (for example “protocol_type” and “group_id”). In most “normal” use cases however, the values of these advanced fields are consistent for each event instance (for example, a single trace contains events for a single QUIC connection). To reduce file size and making logging easier, qlog uses the “common_fields” list to indicate those fields and their values that are shared by all events in this component trace. This prevents these fields from being logged for each individual event. An example of this is shown in .

The “common_fields” field is a generic dictionary of key-value pairs, where the key is always a string and the value can be of any type, but is typically also a string or number. As such, unknown entries in this dictionary MUST be disregarded by the user and tools (i.e., the presence of an uknown field is explicitly NOT an error). The list of default qlog fields that are typically logged in common_fields (as opposed to as individual fields per event instance) are: time_format reference_time protocol_type group_id Tools MUST be able to deal with these fields being defined either on each event individually or combined in common_fields. Note that if at least one event in a trace has a different value for a given field, this field MUST NOT be added to common_fields but instead defined on each event individually. Good example of such fields are “time” and “data”, who are divergent by nature.

This document only defines the main schema for the qlog format. This is intended to be used together with specific, per-protocol event definitions that specify the name (category + type) and data needed for each individual event. This is with the intent to allow the qlog main schema to be easily re-used for several protocols. Examples include the QUIC event definitions and HTTP/3 and QPACK event definitions . This section defines some basic annotations and concepts the creators of event definition documents SHOULD follow to ensure a measure of consistency, making it easier for qlog implementers to extrapolate from one protocol to another.

TODO: pending QUIC working group discussion. This text reflects the initial (qlog draft 01 and 02) setup. There are several ways of defining qlog events. In practice, we have seen two main types used so far: a) those that map directly to concepts seen in the protocols (e.g., packet_sent) and b) those that act as aggregating events that combine data from several possible protocol behaviours or code paths into one (e.g., parameters_set). The latter are typically used as a means to reduce the amount of unique event definitions, as reflecting each possible protocol event as a separate qlog entity would cause an explosion of event types. Additionally, logging duplicate data is typically prevented as much as possible. For example, packet header values that remain consistent across many packets are split into separate events (for example spin_bit_updated or connection_id_updated for QUIC). Finally, we have typically refrained from adding additional state change events if those state changes can be directly inferred from data on the wire (for example flow control limit changes) if the implementation is bug-free and spec-compliant. Exceptions have been made for common events that benefit from being easily identifiable or individually logged (for example packets_acked).

Depending on how events are designed, it may be that several events allow the logging of similar or overlapping data. For example the separate QUIC connection_started event overlaps with the more generic connection_state_updated. In these cases, it is not always clear which event should be logged or used, and which event should take precedence if e.g., both are present and provide conflicting information. To aid in this decision making, we recommend that each event SHOULD have an “importance indicator” with one of three values, in decreasing order of importance and exptected usage: Core Base Extra The “Core” events are the events that SHOULD be present in all qlog files for a given protocol. These are typically tied to basic packet and frame parsing and creation, as well as listing basic internal metrics. Tool implementers SHOULD expect and add support for these events, though SHOULD NOT expect all Core events to be present in each qlog trace. The “Base” events add additional debugging options and CAN be present in qlog files. Most of these can be implicitly inferred from data in Core events (if those contain all their properties), but for many it is better to log the events explicitly as well, making it clearer how the implementation behaves. These events are for example tied to passing data around in buffers, to how internal state machines change and help show when decisions are actually made based on received data. Tool implementers SHOULD at least add support for showing the contents of these events, if they do not handle them explicitly. The “Extra” events are considered mostly useful for low-level debugging of the implementation, rather than the protocol. They allow more fine-grained tracking of internal behaviour. As such, they CAN be present in qlog files and tool implementers CAN add support for these, but they are not required to. Note that in some cases, implementers might not want to log for example data content details in the “Core” events due to performance or privacy considerations. In this case, they SHOULD use (a subset of) relevant “Base” events instead to ensure usability of the qlog output. As an example, implementations that do not log QUIC packet_received events and thus also not which (if any) ACK frames the packet contains, SHOULD log packets_acked events instead. Finally, for event types whose data (partially) overlap with other event types’ definitions, where necessary the event definition document should include explicit guidance on which to use in specific situations.

Event definition documents are free to define new category and event types, top-level fields (e.g., a per-event field indicating its privacy properties or path_id in multipath protocols), as well as values for the “trigger” property within the “data” field, or other member fields of the “data” field, as they see fit. They however SHOULD NOT expect non-specialized tools to recognize or visualize this custom data. However, tools SHOULD make an effort to visualize even unknown data if possible in the specific tool’s context. If they do not, they MUST ignore these unknown fields.

There are some event types and data classes that are common across protocols, applications and use cases that benefit from being defined in a single location. This section specifies such common definitions.

While qlog is a more high-level logging format, it also allows the inclusion of most raw wire image information, such as byte lengths and even raw byte values. This can be useful when for example investigating or tuning packetization behaviour or determining encoding/framing overheads. However, these fields are not always necessary and can take up considerable space if logged for each packet or frame. They can also have a considerable privacy and security impact. As such, they are grouped in a separate optional field called “raw” of type RawInfo (where applicable).

The RawInfo:data field can be truncated for privacy or security purposes (for example excluding payload data). In this case, the length properties should still indicate the non-truncated lengths. We do not specify explicit header_length or trailer_length fields. In most protocols, header_length can be calculated by subtracing the payload_length from the length (e.g., if trailer_length is always 0). In protocols with trailers (e.g., QUIC’s AEAD tag), event definitions documents SHOULD define other ways of logging the trailer_length to make the header_length calculation possible. The exact definitions entities, headers, trailers and payloads depend on the protocol used. If this is non-trivial, event definitions documents SHOULD include a clear explanation of how entities are mapped into the RawInfo structure. Relatedly, many modern protocols use Variable-Length Integer Encoded (VLIE) values in their headers, which are of a dynamic length. Because of this, we cannot deterministally reconstruct the header encoding/length from non-RawInfo qlog data, as implementations might not necessarily employ the most efficient VLIE scheme for all values. As such, to make exact size-analysis possible, implementers should use explicit lengths in RawInfo rather than reconstructing them from other qlog data. Similarly, tool developers should only utilize RawInfo (and related information) in such tools to prevent errors.

In typical logging setups, users utilize a discrete number of well-defined logging categories, levels or severities to log freeform (string) data. This generic events category replicates this approach to allow implementations to fully replace their existing text-based logging by qlog. This is done by providing events to log generic strings for the typical well-known logging levels (error, warning, info, debug, verbose). For the events defined below, the “category” is “generic” and their “type” is the name of the heading in lowercase (e.g., the “name” of the error event is “generic:error”).

Importance: Core Used to log details of an internal error that might not get reflected on the wire. Data:

Importance: Base Used to log details of an internal warning that might not get reflected on the wire. Data:

Importance: Extra Used mainly for implementations that want to use qlog as their one and only logging format but still want to support unstructured string messages. Data:

Importance: Extra Used mainly for implementations that want to use qlog as their one and only logging format but still want to support unstructured string messages. Data:

Importance: Extra Used mainly for implementations that want to use qlog as their one and only logging format but still want to support unstructured string messages. Data:

When evaluating a protocol implementation, one typically sets up a series of interoperability or benchmarking tests, in which the test situations can change over time. For example, the network bandwidth or latency can vary during the test, or the network can be fully disable for a short time. In these setups, it is useful to know when exactly these conditions are triggered, to allow for proper correlation with other events. For the events defined below, the “category” is “simulation” and their “type” is the name of the heading in lowercase (e.g., the “name” of the scenario event is “simulation:scenario”).

Importance: Extra Used to specify which specific scenario is being tested at this particular instance. This could also be reflected in the top-level qlog’s summary or configuration fields, but having a separate event allows easier aggregation of several simulations into one trace (e.g., split by group_id).

Importance: Extra Used to indicate when specific emulation conditions are triggered at set times (e.g., at 3 seconds in 2% packet loss is introduced, at 10s a NAT rebind is triggered).

This document and other related qlog schema definitions are intentionally serialization-format agnostic. This means that implementers themselves can choose how to represent and serialize qlog data practically on disk or on the wire. Some examples of possible formats are JSON, CBOR, CSV, protocol buffers, flatbuffers, etc. All these formats make certain tradeoffs between flexibility and efficiency, with textual formats like JSON typically being more flexible but also less efficient than binary formats like protocol buffers. The format choice will depend on the practical use case of the qlog user. For example, for use in day to day debugging, a plaintext readable (yet relatively large) format like JSON is probably preferred. However, for use in production, a more optimized yet restricted format can be better. In this latter case, it will be more difficult to achieve interoperability between qlog implementations of various protocol stacks, as some custom or tweaked events from one might not be compatible with the format of the other. This will also reflect in tooling: not all tools will support all formats. This being said, the authors prefer JSON as the basis for storing qlog, as it retains full flexibility and maximum interoperability. Storage overhead can be managed well in practice by employing compression. For this reason, this document details both how to practically transform qlog schema definitions to JSON and to the streamable NDJSON. We discuss concrete options to bring down JSON size and processing overheads in . As depending on the employed format different deserializers/parsers should be used, the “qlog_format” field is used to indicate the chosen serialization approach. This field is always a string, but can be made hierarchical by the use of the “.” separator between entries. For example, a value of “JSON.optimizationA” can indicate that a default JSON format is being used, but that a certain optimization of type A was applied to the file as well (see also ).

When mapping qlog to normal JSON, the “qlog_format” field MUST have the value “JSON”. This is also the default qlog serialization and default value of this field. To facilitate this mapping, the qlog documents employ a format that is close to pure JSON for its examples and data definitions. Still, as JSON is not a typed format, there are some practical peculiarities to observe.

While JSON has built-in support for integers up to 64 bits in size, not all JSON parsers do. For example, none of the major Web browsers support full 64-bit integers at this time, as all numerical values (both floating-point numbers and integers) are internally represented as floating point IEEE 754 values. In practice, this limits their integers to a maximum value of 2^53-1. Integers larger than that are either truncated or produce a JSON parsing error. While this is expected to improve in the future (as “BigInt” support has been introduced in most Browsers, though not yet integrated into JSON parsers), we still need to deal with it here. When transforming an int64, uint64 or double from qlog to JSON, the implementer can thus choose to either log them as JSON numbers (taking the risk of truncation or un-parseability) or to log them as strings instead. Logging as strings should however only be practically needed if the value is likely to exceed 2^53-1. In practice, even though protocols such as QUIC allow 64-bit values for for example stream identifiers, these high numbers are unlikely to be reached for the overwhelming majority of cases. As such, it is probably a valid trade-off to take the risk and log 64-bit values as JSON numbers instead of strings. Tools processing JSON-based qlog SHOULD however be able to deal with 64-bit fields being serialized as either strings or numbers.

Unlike most binary formats, JSON does not allow the logging of raw binary blobs directly. As such, when serializing a byte or array<byte>, a scheme needs to be chosen. To represent qlog bytes in JSON, they MUST be serialized to their lowercase hexadecimal equivalents (with 0 prefix for values lower than 10). All values are directly appended to each other, without delimiters. The full value is not prefixed with 0x (as is sometimes common). An example is given in .

As such, the resulting string will always have an even amount of characters and the original byte-size can be retrieved by dividing the string length by 2.

In some cases, it can be interesting not to log a full raw blob but instead a truncated value (for example, only the first 100 bytes of an HTTP response body to be able to discern which file it actually contained). In these cases, the original byte-size length cannot be obtained from the serialized value directly. As such, all qlog schema definitions SHOULD include a separate, length-indicating field for all fields of type array<byte> they specify. This allows always retrieving the original length, but also allows the omission of any raw value bytes of the field completely (e.g., out of privacy or security considerations). To reduce overhead however and in the case the full raw value is logged, the extra length-indicating field can be left out. As such, tools MUST be able to deal with this situation and derive the length of the field from the raw value if no separate length-indicating field is present. All possible permutations are shown by example in .

By definition, JSON strings are serialized surrounded by quotes. Numbers without. qlog type JSON type int8 number int16 number int32 number uint8 number uint16 number uint32 number float number int64 number or string uint64 number or string double number or string bytes string (lowercase hex value) string string boolean string (“true” or “false”) enum string (full value/name, not index) any object ( {…} ) array array ( […] )

JSON files by definition () MUST utilize the UTF-8 encoding, both for the file itself and the string values. Most JSON parsers strictly follow the JSON specification. This includes the rule that trailing comma’s are not allowed. As it is frequently annoying to remove these trailing comma’s when logging events in a streaming fashion, tool implementers SHOULD allow the last event entry of a qlog trace to be an empty object. This allows loggers to simply close the qlog file by appending “{}]}]}” after their last added event. Finally, while not specifically required by the JSON specification, all qlog field names in a JSON serialization MUST be lowercase.

One of the downsides of using pure JSON is that it is inherently a non-streamable format. Put differently, it is not possible to simply append new qlog events to a log file without “closing” this file at the end by appending “]}]}”. Without these closing tags, most JSON parsers will be unable to parse the file entirely. As most platforms do not provide a standard streaming JSON parser (which would be able to deal with this problem), this document also provides a qlog mapping to a streamable JSON format called Newline-Delimited JSON (NDJSON). When mapping qlog to NDJSON, the “qlog_format” field MUST have the value “NDJSON”. NDJSON is very similar to JSON, except that it interprets each line in a file as a fully separate JSON object. Put differently, unlike default JSON, it does not require a file to be wrapped as a full object with “{ … }” or “[ … ]”. Using this setup, qlog events can simply be appended as individually serialized lines at the back of a streamed logging file. For this to work, some qlog definitions have to be adjusted however. Mainly, events are no longer part of the “events” array in the Trace object, but are instead logged separately from the qlog “file header” (QlogFile class in ). Additionally, qlog’s NDJSON mapping does not allow logging multiple individual traces in a single qlog file. As such, the QlogFile:traces field is replaced by the singular “trace” field, which simply contains the Trace data directly. An example can be seen in . Note that the “group_id” field can still be used on a per-event basis to include events from conceptually different sources in a single NDJSON qlog file. Note as well from that the file’s header (QlogFileNDJSON) also needs to be fully serialized on a single line to be NDJSON compatible.

Finally, while not specifically required by the NDJSON specification, all qlog field names in a NDJSON serialization MUST be lowercase.

Note that NDJSON is not supported in most default programming environments (unlike normal JSON). However, several custom NDJSON parsing libraries exist that can be used and the format is easy enough to parse with existing implementations (i.e., by splitting the file into its component lines and feeding them to a normal JSON parser individually, as each line by itself is a valid JSON object).

Both the JSON and NDJSON formatting options described above are serviceable in general small to medium scale (debugging) setups. However, these approaches tend to be relatively verbose, leading to larger file sizes. Additionally, generalized (ND)JSON (de)serialization performance is typically (slightly) lower than that of more optimized and predictable formats. Both aspects make these formats more challenging (though still practical) to use in large scale setups. During the development of qlog, we compared a multitude of alternative formatting and optimization options. The results of this study are summarized on the qlog github repository. The rest of this section discusses some of these approaches implementations could choose and the expected gains and tradeoffs inherent therein. Tools SHOULD support mainly the compression options listed in , as they provide the largest wins for the least cost overall. Over time, specific qlog formats and encodings can be created that more formally define and combine some of the discussed optimizations or add new ones. We choose to define these schemes in separate documents to keep the main qlog definition clean and generalizable, as not all contexts require the same performance or flexibility as others and qlog is intended to be a broadly usable and extensible format (for example more flexibility is needed in earlier stages of protocol development, while more performance is typically needed in later stages). This is also the main reason why the general qlog format is the less optimized JSON instead of a more performant option. To be able to easily distinguish between these options in qlog compatible tooling (without the need to have the user provide out-of-band information or to (heuristically) parse and process files in a multitude of ways, see also ), we recommend using explicit file extensions to indicate specific formats. As there are no standards in place for this type of extension to format mapping, we employ a commonly used scheme here. Our approach is to list the applied optimizations in the extension in ascending order of application (e.g., if a qlog file is first optimized with technique A and then compressed with technique B, the resulting file would have the extension “.qlog.A.B”). This allows tooling to start at the back of the extension to “undo” applied optimizations to finally arrive at the expected qlog representation.

The first general category of optimizations is to alter the representation of data within an (ND)JSON qlog file to reduce file size. The first option is to employ a scheme similar to the CSV (comma separated value ) format, which utilizes the concept of column “headers” to prevent repeating field names for each datapoint instance. Concretely for JSON qlog, several field names are repeated with each event (i.e., time, name, data). These names could be extracted into a separate list, after which qlog events could be serialized as an array of values, as opposed to a full object. This approach was a key part of the original qlog format (prior to draft 02) using the “event_fields” field. However, tests showed that this optimization only provided a mean file size reduction of 5% (100MB to 95MB) while significantly increasing the implementation complexity, and this approach was abandoned in favor of the default JSON setup. Implementations using this format should not employ a separate file extension (as it still uses JSON), but rather employ a new value of “JSON.namedheaders” (or “NDJSON.namedheaders”) for the “qlog_format” field (see ). The second option is to replace field values and/or names with indices into a (dynamic) lookup table. This is a common compression technique and can provide significant file size reductions (up to 50% in our tests, 100MB to 50MB). However, this approach is even more difficult to implement efficiently and requires either including the (dynamic) table in the resulting file (an approach taken by for example Chromium’s NetLog format) or defining a (static) table up-front and sharing this between implementations. Implementations using this approach should not employ a separate file extension (as it still uses JSON), but rather employ a new value of “JSON.dictionary” (or “NDJSON.dictionary”) for the “qlog_format” field (see ). As both options either proved difficult to implement, reduced qlog file readability, and provided too little improvement compared to other more straightforward options (for example ), these schemes are not inherently part of qlog.

The second general category of optimizations is to utilize a (generic) compression scheme for textual data. As qlog in the (ND)JSON format typically contains a large amount of repetition, off-the-shelf (text) compression techniques typically succeed very well in bringing down file sizes (regularly with up to two orders of magnitude in our tests, even for “fast” compression levels). As such, utilizing compression is recommended before attempting other optimization options, even though this might (somewhat) increase processing costs due to the additional compression step. The first option is to use GZIP compression (). This generic compression scheme provides multiple compression levels (providing a trade-off between compression speed and size reduction). Utilized at level 6 (a medium setting thought to be applicable for streaming compression of a qlog stream in commodity devices), gzip compresses qlog JSON files to 7% of their initial size on average (100MB to 7MB). For this option, the file extension .qlog.gz SHOULD BE used. The “qlog_format” field should still reflect the original JSON formatting of the qlog data (e.g., “JSON” or “NDJSON”). The second option is to use Brotli compression (). While similar to gzip, this more recent compression scheme provides a better efficiency. It also allows multiple compression levels. Utilized at level 4 (a medium setting thought to be applicable for streaming compression of a qlog stream in commodity devices), brotli compresses qlog JSON files to 7% of their initial size on average (100MB to 7MB). For this option, the file extension .qlog.br SHOULD BE used. The “qlog_format” field should still reflect the original JSON formatting of the qlog data (e.g., “JSON” or “NDJSON”). Other compression algorithms of course exist (for example xz, zstd, and lz4). We mainly recommend gzip and brotli because of their tweakable behaviour and wide support in web-based environments, which we envision as the main tooling ecosystem (see also ).

The third general category of optimizations is to use a more optimized (often binary) format instead of the textual JSON format. This approach inherently produces smaller files and often has better (de)serialization performance. However, the resultant files are no longer human readable and some formats require hard tradeoffs between flexibility for performance. The first option is to use the CBOR (Concise Binary Object Representation ) format. For our purposes, CBOR can be viewed as a straighforward binary variant of JSON. As such, existing JSON qlog files can be trivially converted to and from CBOR (though slightly more work is needed for NDJSON qlogs). While CBOR thus does retain the full qlog flexibility, it only provides a 25% file size reduction (100MB to 75MB) compared to textual (ND)JSON. As CBOR support in programming environments is not as widespread as that of textual JSON and the format lacks human readability, CBOR was not chosen as the default qlog format. For this option, the file extension .qlog.cbor SHOULD BE used. The “qlog_format” field should still reflect the original JSON formatting of the qlog data (e.g., “JSON” or “NDJSON”). A second option is to use a more specialized binary format, such as Protocol Buffers (protobuf). This format is battle-tested, has support for optional fields and has libraries in most programming languages. Still, it is significantly less flexible than textual JSON or CBOR, as it relies on a separate, pre-defined schema (a .proto file). As such, it it not possible to (easily) log new event types in protobuf files without adjusting this schema as well, which has its own practical challenges. As qlog is intended to be a flexible, general purpose format, this type of format was not chosen as its basic serialization. The lower flexibility does lead to significantly reduced file sizes. Our straightforward mapping of the qlog main schema and QUIC/HTTP3 event types to protobuf created qlog files 24% as large as the raw JSON equivalents (100MB to 24MB). For this option, the file extension .qlog.protobuf SHOULD BE used. The “qlog_format” field should reflect the different internal format, for example: “qlog_format”: “protobuf”. Note that binary formats can (and should) also be used in conjunction with compression (see ). For example, CBOR compresses well (to about 6% of the original textual JSON size (100MB to 6MB) for both gzip and brotli) and so does protobuf (5% (gzip) to 3% (brotli)). However, these gains are similar to the ones achieved by simply compression the textual JSON equivalents directly (7%, see ). As such, since compression is still needed to achieve optimal file size reductions event with binary formats, we feel the more flexible compressed textual JSON options are a better default for the qlog format in general.

In summary, textual JSON was chosen as the main qlog format due to its high flexibility and because its inefficiencies can be largely solved by the utilization of compression techniques (which are needed to achieve optimal results with other formats as well). Still, qlog implementers are free to define other qlog formats depending on their needs and context of use. These formats should be described in their own documents, the discussion in this document mainly acting as inspiration and high-level guidance. Implementers are encouraged to add concrete qlog formats and definitions to the designated public repository. The following table provides an overview of all the discussed qlog formatting options with examples: format qlog_format extension JSON JSON .qlog NDJSON NDJSON .qlog named headers (ND)JSON.namedheaders .qlog dictionary (ND)JSON.dictionary .qlog CBOR (ND)JSON .qlog.cbor protobuf protobuf .qlog.protobuf       gzip no change .gz suffix brotli no change .br suffix

As discussed in the previous sections, a qlog file can be serialized in a multitude of formats, each of which can conceivably be transformed into or from one another without loss of information. For example, a number of NDJSON streamed qlogs could be combined into a JSON formatted qlog for later processing. Similarly, a captured binary qlog could be transformed to JSON for easier interpretation and sharing. Secondly, we can also consider other structured logging approaches that contain similar (though typically not identical) data to qlog, like raw packet capture files (for example .pcap files from tcpdump) or endpoint-specific logging formats (for example the NetLog format in Google Chrome). These are sometimes the only options, if an implementation cannot or will not support direct qlog output for any reason, but does provide other internal or external (e.g., SSLKEYLOGFILE export to allow decryption of packet captures) logging options For this second category, a (partial) transformation from/to qlog can also be defined. As such, when defining a new qlog serialization format or wanting to utilize qlog-compatible tools with existing codebases lacking qlog support, it is recommended to define and provide a concrete mapping from one format to default JSON-serialized qlog. Several of such mappings exist. Firstly, [pcap2qlog]((https://github.com/quiclog/pcap2qlog) transforms QUIC and HTTP/3 packet capture files to qlog. Secondly, netlog2qlog converts chromium’s internal dictionary-encoded JSON format to qlog. Finally, quictrace2qlog converts the older quictrace format to JSON qlog. Tools can then easily integrate with these converters (either by incorporating them directly or for example using them as a (web-based) API) so users can provide different file types with ease. For example, the qvis toolsuite supports a multitude of formats and qlog serializations.

Different implementations will have different ways of generating and storing qlogs. However, there is still value in defining a few default ways in which to steer this generation and access of the results.

To provide users control over where and how qlog files are created, we define two environment variables. The first, QLOGFILE, indicates a full path to where an individual qlog file should be stored. This path MUST include the full file extension. The second, QLOGDIR, sets a general directory path in which qlog files should be placed. This path MUST include the directory separator character at the end. In general, QLOGDIR should be preferred over QLOGFILE if an endpoint is prone to generate multiple qlog files. This can for example be the case for a QUIC server implementation that logs each QUIC connection in a separate qlog file. An alternative that uses QLOGFILE would be a QUIC server that logs all connections in a single file and uses the “group_id” field () to allow post-hoc separation of events. Implementations SHOULD provide support for QLOGDIR and MAY provide support for QLOGFILE. When using QLOGDIR, it is up to the implementation to choose an appropriate naming scheme for the qlog files themselves. The chosen scheme will typically depend on the context or protocols used. For example, for QUIC, it is recommended to use the Original Destination Connection ID (ODCID), followed by the vantage point type of the logging endpoint. Examples of all options for QUIC are shown in .

After generation, qlog implementers MAY make available generated logs and traces on an endpoint (typically the server) via the following .well-known URI: .well-known/qlog/IDENTIFIER.extension The IDENTIFIER variable depends on the context and the protocol. For example for QUIC, the lowercase Original Destination Connection ID (ODCID) is recommended, as it can uniquely identify a connection. Additionally, the extension depends on the chosen format (see ). For example, for a QUIC connection with ODCID “abcde”, the endpoint for fetching its default JSON-formatted .qlog file would be: .well-known/qlog/abcde.qlog Implementers SHOULD allow users to fetch logs for a given connection on a 2nd, separate connection. This helps prevent pollution of the logs by fetching them over the same connection that one wishes to observe through the log. Ideally, for the QUIC use case, the logs should also be approachable via an HTTP/2 or HTTP/1.1 endpoint (i.e., on TCP port 443), to for example aid debugging in the case where QUIC/UDP is blocked on the network. qlog implementers SHOULD NOT enable this .well-known endpoint in typical production settings to prevent (malicious) users from downloading logs from other connections. Implementers are advised to disable this endpoint by default and require specific actions from the end users to enable it (and potentially qlog itself). Implementers MUST also take into account the general privacy and security guidelines discussed in before exposing qlogs to outside actors.

Tools ingestion qlog MUST indicate which qlog version(s), qlog format(s), compression methods and potentially other input file formats (for example .pcap) they support. Tools SHOULD at least support .qlog files in the default JSON format (). Additionally, they SHOULD indicate exactly which values for and properties of the name (category and type) and data fields they look for to execute their logic. Tools SHOULD perform a (high-level) check if an input qlog file adheres to the expected qlog schema. If a tool determines a qlog file does not contain enough supported information to correctly execute the tool’s logic, it SHOULD generate a clear error message to this effect. Tools MUST NOT produce breaking errors for any field names and/or values in the qlog format that they do not recognize. Tools SHOULD indicate even unknown event occurences within their context (e.g., marking unknown events on a timeline for manual interpretation by the user). Tool authors should be aware that, depending on the logging implementation, some events will not always be present in all traces. For example, using a circular logging buffer of a fixed size, it could be that the earliest events (e.g., connection setup events) are later overwritten by “newer” events. Alternatively, some events can be intentionally omitted out of privacy or file size considerations. Tool authors are encouraged to make their tools robust enough to still provide adequate output for incomplete logs.

TODO : discuss privacy and security considerations (e.g., what NOT to log, what to strip out of a log before sharing, …) TODO: strip out/don’t log IPs, ports, specific CIDs, raw user data, exact times, HTTP HEADERS (or at least :path), SNI values TODO: see if there is merit in encrypting the logs and having the server choose an encryption key (e.g., sent in transport parameters) Good initial reference: Christian Huitema’s blogpost

TODO: primarily the .well-known URI

KU Leuven Facebook Protocol Labs KU Leuven Facebook Protocol Labs In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. JavaScript Object Notation (JSON) is a lightweight, text-based, language-independent data interchange format. It was derived from the ECMAScript Programming Language Standard. JSON defines a small set of formatting rules for the portable representation of structured data.This document removes inconsistencies with other specifications of JSON, repairs specification errors, and offers experience-based interoperability guidance. This RFC documents the format used for Comma-Separated Values (CSV) files and registers the associated MIME type "text/csv". This memo provides information for the Internet community. This specification defines a lossless compressed data format that is compatible with the widely used GZIP utility. This memo provides information for the Internet community. This memo does not specify an Internet standard of any kind. This specification defines a lossless compressed data format that compresses data using a combination of the LZ77 algorithm and Huffman coding, with efficiency comparable to the best currently available general-purpose compression methods. The Concise Binary Object Representation (CBOR) is a data format whose design goals include the possibility of extremely small code size, fairly small message size, and extensibility without the need for version negotiation. These design goals make it different from earlier binary serializations such as ASN.1 and MessagePack.

These changes were done in preparation of the adoption of the drafts by the QUIC working group (#137) Moved RawInfo, Importance, Generic events and Simulation events to this document. Added basic event definition guidelines Made protocol_type an array instead of a string (#146)

Decoupled qlog from the JSON format and described a mapping instead (#89) Data types are now specified in this document and proper definitions for fields were added in this format 64-bit numbers can now be either strings or numbers, with a preference for numbers (#10) binary blobs are now logged as lowercase hex strings (#39, #36) added guidance to add length-specifiers for binary blobs (#102) Removed “time_units” from Configuration. All times are now in ms instead (#95) Removed the “event_fields” setup for a more straightforward JSON format (#101,#89) Added a streaming option using the NDJSON format (#109,#2,#106) Described optional optimization options for implementers (#30) Added QLOGDIR and QLOGFILE environment variables, clarified the .well-known URL usage (#26,#33,#51) Overall tightened up the text and added more examples

All field names are now lowercase (e.g., category instead of CATEGORY) Triggers are now properties on the “data” field value, instead of separate field types (#23) group_ids in common_fields is now just also group_id

Quic-trace takes a slightly different approach based on protocolbuffers. Spindump also defines a custom text-based format for in-network measurements Wireshark also has a QUIC dissector and its results can be transformed into a json output format using tshark. The idea is that qlog is able to encompass the use cases for both of these alternate designs and that all tooling converges on the qlog standard.

Much of the initial work by Robin Marx was done at Hasselt University. Thanks to Jana Iyengar, Brian Trammell, Dmitri Tikhonov, Stephen Petrides, Jari Arkko, Marcus Ihlar, Victor Vasiliev, Mirja Kühlewind, Jeremy Lainé and Lucas Pardue for their feedback and suggestions.