[llvm] [llvm]Add a simple Telemetry framework (PR #102323)

James Henderson via llvm-commits llvm-commits at lists.llvm.org
Tue Dec 17 01:40:39 PST 2024


================
@@ -0,0 +1,246 @@
+===========================
+Telemetry framework in LLVM
+===========================
+
+.. contents::
+   :local:
+
+.. toctree::
+   :hidden:
+
+===========================
+Telemetry framework in LLVM
+===========================
+
+Objective
+=========
+
+Provides a common framework in LLVM for collecting various usage and performance
+metrics.
+It is located at ``llvm/Telemetry/Telemetry.h``
+
+Characteristics
+---------------
+* Configurable and extensible by:
+
+  * Tools: any tool that wants to use Telemetry can extend and customize it.
+  * Vendors: Toolchain vendors can also provide custom implementation of the
+    library, which could either override or extend the given tool's upstream
+    implementation, to best fit their organization's usage and privacy models.
+  * End users of such tool can also configure Telemetry (as allowed by their
+    vendor).
+
+Important notes
+----------------
+
+* There is no concrete implementation of a Telemetry library in upstream LLVM.
+  We only provide the abstract API here. Any tool that wants telemetry will
+  implement one.
+  
+  The rationale for this is that all the tools in LLVM are very different in
+  what they care about (what/where/when to instrument data). Hence, it might not
+  be practical to have a single implementation.
+  However, in the future, if we see enough common pattern, we can extract them
+  into a shared place. This is TBD - contributions are welcomed.
+
+* No implementation of Telemetry in upstream LLVM shall store any of the
+  collected data due to privacy and security reasons:
+  
+  * Different organizations have different privacy models:
+  
+    * Which data is sensitive, which is not?
+    * Whether it is acceptable for instrumented data to be stored anywhere?
+      (to a local file, what not?)
+      
+  * Data ownership and data collection consents are hard to accommodate from
+    LLVM developers' point of view:
+  
+    * E.g., data collected by Telemetry is not necessarily owned by the user
+      of an LLVM tool with Telemetry enabled, hence the user's consent to data
+      collection is not meaningful. On the other hand, LLVM developers have no
+      reasonable ways to request consent from the "real" owners.
+
+
+High-level design
+=================
+
+Key components
+--------------
+
+The framework consists of four important classes:
+
+* ``llvm::telemetry::Manager``: The class responsible for collecting and
+  transmitting telemetry data. This is the main point of interaction between the
+  framework and any tool that wants to enable telemetry.
+* ``llvm::telemetry::TelemetryInfo``: Data courier
+* ``llvm::telemetry::Destination``: Data sink to which the Telemetry framework
+  sends data.
+  Its implementation is transparent to the framework.
+  It is up to the vendor to decide which pieces of data to forward and where
+  to forward them to for their final storage.
+* ``llvm::telemetry::Config``: Configurations for the ``Manager``.
+  
+.. image:: llvm_telemetry_design.png
+
+How to implement and interact with the API
+------------------------------------------
+
+To use Telemetry in your tool, you need to provide a concrete implementation of the ``Manager`` class and ``Destination``.
+
+1) Define a custom ``Serializer``, ``Manager``, ``Destination`` and optionally a subclass of ``TelemetryInfo``
+
+.. code-block:: c++
+
+  class JsonSerializer : public Serializer {
+  public:
+    json::Object *getOutputObject() { return object.get(); }
+
+    llvm::Error init() override {
----------------
jh7370 wrote:

Ping this comment?

https://github.com/llvm/llvm-project/pull/102323


More information about the llvm-commits mailing list