[MP][Observability] Add telemetry subsystem for multiprocess mode#2696
[MP][Observability] Add telemetry subsystem for multiprocess mode#2696ApostaC merged 10 commits intoLMCache:devfrom
Conversation
Signed-off-by: ApostaC <yihua98@uchicago.edu>
Signed-off-by: ApostaC <yihua98@uchicago.edu>
Summary of ChangesHello, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request introduces a robust, event-based telemetry system for LMCache's multiprocess mode, significantly enhancing observability by providing detailed, per-operation tracing. The framework is designed for extensibility, allowing for various processors to consume events, and is integrated seamlessly into the server's lifecycle with minimal performance impact when disabled. This foundational work enables deeper insights into cache behavior and performance characteristics. Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Changelog
Activity
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
There was a problem hiding this comment.
Code Review
This pull request introduces a well-designed and comprehensive telemetry subsystem. The architecture is robust, featuring a lock-free queue, a daemon drain thread for processing, and a flexible processor plugin system. The implementation is clean and includes thorough unit tests. I have a couple of minor suggestions for improvement regarding configuration handling to enhance robustness and code clarity.
Signed-off-by: ApostaC <yihua98@uchicago.edu>
|
Example usage command (log the telemetry to terminal) python3 -m lmcache.v1.multiprocess.server --max-workers 4 --port 6555 \
--l1-size 70 --eviction-policy LRU \
--l2-adapter '{"type": "mock", "max_size_gb": 200, "mock_bandwidth_gb": 10}' \
--enable-telemetry --telemetry-processor '{"type": "logging", "log_level":"warning"}'Expected output: |
Signed-off-by: ApostaC <yihua98@uchicago.edu>
There was a problem hiding this comment.
Some quick comments:
- previously IIRC you prefer to keep the
lmcache/v1for pure storage rather than request-level telemetry. Is that still true? - Please also re-implement the request_finished telemetry under the integration folder using this framework, and deprecate the one under integration folder.
- When request fully ends, it would be really nice if we can have a "summary" of this request logged out, to avoid people manually align the time. Example:
Summary of request chat-cmpl-xxx:
Received at (time)
Preprocessing: ??ms (st->ed: 2:04.025 -> 2.04.124)
Retrieve ?? tokens: ?? ms (st->ed: ???)
LLM computation runs: ??ms (st->ed: ???)
Store ?? tokens: ?? ms (st->ed: ???)
This might be convenient for debugging.
Signed-off-by: ApostaC <yihua98@uchicago.edu>
This one is actually designed for request-level telemetry. The core design here is to have a very fast, non-blocking API to submit the telemetry event (i.e., push it into a queue), and then there is a background thread processing all the telemetry events.
Unfortunately, this could be pretty difficult, since we need to track the TP information at the LMCache-level, which is quite hard. Additionally, though both modules are named as "telemetry", this one is more for the debuggability and visualization (e.g., OpenTelemetry support)
Will have a follow-up PR for it. There should also be some tools to visualize the spans. |
Signed-off-by: Yihua Cheng <yihua98@uchicago.edu>
Signed-off-by: ApostaC <yihua98@uchicago.edu>
| with self._lock: | ||
| self._processors.append(processor) | ||
|
|
||
| def start(self) -> None: |
There was a problem hiding this comment.
should this be idempotent? or doesn't matter
There was a problem hiding this comment.
Good point! Let me fix it
Signed-off-by: ApostaC <yihua98@uchicago.edu>
…nto local-dev/mp-telemetry
…Cache#2696) * [Add] the telemetry subsystem for MP mode Signed-off-by: ApostaC <yihua98@uchicago.edu> * integrate telemetry system for lookup store and retrieve Signed-off-by: ApostaC <yihua98@uchicago.edu>
…Cache#2696) * [Add] the telemetry subsystem for MP mode Signed-off-by: ApostaC <yihua98@uchicago.edu> * integrate telemetry system for lookup store and retrieve Signed-off-by: ApostaC <yihua98@uchicago.edu> Signed-off-by: shaoxiawjc <wjc2800@163.com>
…Cache#2696) * [Add] the telemetry subsystem for MP mode Signed-off-by: ApostaC <yihua98@uchicago.edu> * integrate telemetry system for lookup store and retrieve Signed-off-by: ApostaC <yihua98@uchicago.edu> Signed-off-by: Aaron Wu <aaron.wu@dell.com>
…Cache#2696) * [Add] the telemetry subsystem for MP mode Signed-off-by: ApostaC <yihua98@uchicago.edu> * integrate telemetry system for lookup store and retrieve Signed-off-by: ApostaC <yihua98@uchicago.edu>
…Cache#2696) * [Add] the telemetry subsystem for MP mode Signed-off-by: ApostaC <yihua98@uchicago.edu> * integrate telemetry system for lookup store and retrieve Signed-off-by: ApostaC <yihua98@uchicago.edu>
What this PR does / why we need it:
Adds an event-based telemetry framework for MP mode observability, complementing the existing Prometheus metrics (aggregated counters/histograms) with per-operation event tracing.
Design: Uses a START/END event model with session IDs. Call sites emit a
STARTevent when an operation begins and anENDevent when it completes, sharing the samesession_idto correlate them into spans. This two-event design is necessary because start and end happen at different call sites in LMCache's async submit/check/complete patterns. SeeDESIGN.mdfor full architecture details and how to implement new processors.Key components:
TelemetryEventdataclass (name, event_type, session_id, metadata)TelemetryProcessor/TelemetryProcessorConfigABCs with a registry pattern (same as L2 adapters)LoggingProcessorbuilt-in processor with configurable log levelTelemetryControllerwith lock-free deque, daemon drain thread, tail-drop backpressuremake_start_event()/make_end_event()convenience functions for concise call sitesUser-facing changes:
--enable-telemetry,--telemetry-max-queue-size,--telemetry-processor JSON--enable-telemetry --telemetry-processor '{"type": "logging", "log_level": "INFO"}'store,retrieve, andlookupoperations inserver.pyserver.pyandblend_server.pystartup/shutdown lifecycleSpecial notes for your reviewers:
lmcache/v1/distributed/l2_adapters/config.pycollections.dequeis used without a lock — CPython GIL guarantees atomicappend()/popleft()with a single consumer threadlaunch_host_funcwithis_enabled()guards to avoid registering unnecessary CUDA callbacks when telemetry is disabledIf applicable: