<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Ti Gg</title>
    <description>The latest articles on DEV Community by Ti Gg (@ti_gg_043b3e06a4b7d5d0888).</description>
    <link>https://dev.to/ti_gg_043b3e06a4b7d5d0888</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4210925%2F1b12cfa9-134c-48d5-9656-b5808f5e21ae.png</url>
      <title>DEV Community: Ti Gg</title>
      <link>https://dev.to/ti_gg_043b3e06a4b7d5d0888</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/ti_gg_043b3e06a4b7d5d0888"/>
    <language>en</language>
    <item>
      <title>Stop Leaving PyCharm to Check a Training Run</title>
      <dc:creator>Ti Gg</dc:creator>
      <pubDate>Mon, 16 Nov 2026 11:26:45 +0000</pubDate>
      <link>https://dev.to/ti_gg_043b3e06a4b7d5d0888/stop-leaving-pycharm-to-check-a-training-run-4kp1</link>
      <guid>https://dev.to/ti_gg_043b3e06a4b7d5d0888/stop-leaving-pycharm-to-check-a-training-run-4kp1</guid>
      <description>&lt;p&gt;A training job finishes. The interesting part starts: which run was that, what were its&lt;br&gt;
parameters, and did that loss curve actually go down? For most ML teams the answer lives in three&lt;br&gt;
browser tabs — the MLflow UI on port 5000, a W&amp;amp;B project, an Optuna dashboard — and every check&lt;br&gt;
costs a context switch, a re-authentication, and a new set of tabs to lose.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://plugins.jetbrains.com/plugin/35416-ml-experiment-cockpit" rel="noopener noreferrer"&gt;ML Experiment Cockpit&lt;/a&gt; is a free&lt;br&gt;
JetBrains plugin that puts a &lt;strong&gt;read-only&lt;/strong&gt; view of those servers inside a tool window in PyCharm,&lt;br&gt;
IntelliJ IDEA, WebStorm, GoLand or CLion. It does one thing: it shows you what your tracking&lt;br&gt;
servers already contain, without leaving the editor.&lt;/p&gt;
&lt;h2&gt;
  
  
  What it shows you
&lt;/h2&gt;

&lt;p&gt;Four tabs, one per kind of server:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Experiments&lt;/strong&gt; — your MLflow experiment tree (experiments, runs, status, parameters). Pick a run,
pick a metric, and the plugin charts its history: the curve, the last value, and the step the
value was recorded at.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Optuna&lt;/strong&gt; — studies and trials read from an Optuna Dashboard.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Models&lt;/strong&gt; — the MLflow Model Registry (registered models and their versions).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sweeps&lt;/strong&gt; — W&amp;amp;B sweeps, read-only, with your own W&amp;amp;B API key.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The last three tabs stay empty until you configure them, so a plain MLflow setup only ever sees&lt;br&gt;
the first tab.&lt;/p&gt;
&lt;h2&gt;
  
  
  Getting started
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Open the tool window: &lt;strong&gt;View | Tool Windows | ML Experiment Cockpit&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Set your server in &lt;strong&gt;Settings | Tools | ML Experiment Cockpit&lt;/strong&gt;: the MLflow tracking URI, for
example &lt;code&gt;http://localhost:5000&lt;/code&gt;. Leave the other fields empty if you do not use Optuna or W&amp;amp;B.&lt;/li&gt;
&lt;li&gt;Press &lt;strong&gt;Refresh&lt;/strong&gt; in the tool window toolbar.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;That is the whole setup. There is no account, no key to obtain for MLflow, and no service of ours&lt;br&gt;
in the middle.&lt;/p&gt;
&lt;h2&gt;
  
  
  The code you would chart
&lt;/h2&gt;

&lt;p&gt;The plugin never runs your training code and never writes to your tracking server — it reads what&lt;br&gt;
the server already has, the same way the MLflow SDK does. So the code stays exactly where it&lt;br&gt;
belongs, in your project:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="n"&gt;mlflow&lt;/span&gt;

&lt;span class="n"&gt;mlflow&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;set_tracking_uri&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"http://localhost:5000"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;with&lt;/span&gt; &lt;span class="n"&gt;mlflow&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;start_run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;run_name&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"exp-42"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;mlflow&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;log_param&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"lr"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mf"&gt;3e-4&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;mlflow&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;log_param&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"batch_size"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;32&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;step&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="nb"&gt;range&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;200&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;mlflow&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;log_metric&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"loss"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mf"&gt;1.0&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;step&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;step&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;step&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Start that, then refresh the tool window: &lt;code&gt;exp-42&lt;/code&gt; appears under your experiment with &lt;code&gt;lr&lt;/code&gt; and&lt;br&gt;
&lt;code&gt;batch_size&lt;/code&gt; in the parameters panel, and the &lt;code&gt;loss&lt;/code&gt; metric is one click away as a curve. If the&lt;br&gt;
run looks wrong, you know which parameter to change before you queue the next one.&lt;/p&gt;

&lt;h2&gt;
  
  
  The limits, so you do not have to discover them
&lt;/h2&gt;

&lt;p&gt;This is a viewer, not a second MLflow client, and it has boundaries on purpose:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;up to &lt;strong&gt;500 experiments&lt;/strong&gt;, &lt;strong&gt;100 runs per experiment&lt;/strong&gt;, &lt;strong&gt;500 registered models&lt;/strong&gt;, &lt;strong&gt;500 model
versions&lt;/strong&gt;, and &lt;strong&gt;5 000 metric points per run&lt;/strong&gt;;&lt;/li&gt;
&lt;li&gt;W&amp;amp;B sweeps are read without pagination — you see the &lt;strong&gt;first 50 sweeps&lt;/strong&gt;;&lt;/li&gt;
&lt;li&gt;it is &lt;strong&gt;read-only&lt;/strong&gt;: it will not start a run, change a parameter, tag a model or delete anything;&lt;/li&gt;
&lt;li&gt;it renders what your server returns within those limits, so if a server is configured to hide
something, the plugin cannot show it.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you need to &lt;em&gt;write&lt;/em&gt; to MLflow from a script, keep using the MLflow SDK — that is still the&lt;br&gt;
right tool.&lt;/p&gt;

&lt;h2&gt;
  
  
  What happens to your data
&lt;/h2&gt;

&lt;p&gt;The plugin talks to the servers &lt;strong&gt;you&lt;/strong&gt; configure — the MLflow tracking server, the Optuna&lt;br&gt;
dashboard, the W&amp;amp;B endpoint — and to nothing else. There is no telemetry, no analytics, no&lt;br&gt;
advertising, and no code that sends anything to JetBrains or to us. We do not operate a proxy: the&lt;br&gt;
requests go straight from your IDE to your server.&lt;/p&gt;

&lt;p&gt;Your server URLs and W&amp;amp;B entity/project are stored in the IDE's own configuration directory, in&lt;br&gt;
&lt;code&gt;mlExperimentCockpit.xml&lt;/code&gt;, and are &lt;strong&gt;not&lt;/strong&gt; synchronized through IDE Settings Sync. Your W&amp;amp;B API key&lt;br&gt;
lives in the IDE's credential store (PasswordSafe), never in a settings file. If you uninstall&lt;br&gt;
the plugin, the &lt;code&gt;mlExperimentCockpit.xml&lt;/code&gt; file stays behind in the IDE's configuration&lt;br&gt;
directory — your MLflow, Optuna and W&amp;amp;B data stays on your servers either way.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who should skip it
&lt;/h2&gt;

&lt;p&gt;If your workflow is "one notebook, one local run, no tracking server", a plugin is overkill. If you&lt;br&gt;
need to launch training jobs from the IDE, this is not that tool either. If you do keep an MLflow,&lt;br&gt;
Optuna or W&amp;amp;B server, and you keep coming back to it to check a curve, this is the shortcut.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Plugin page (free, published on the JetBrains Marketplace):
&lt;a href="https://plugins.jetbrains.com/plugin/35416-ml-experiment-cockpit" rel="noopener noreferrer"&gt;https://plugins.jetbrains.com/plugin/35416-ml-experiment-cockpit&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Source code (Apache License 2.0):
&lt;a href="https://github.com/tgg49099-dotcom/atlassian-apps/tree/main/jetbrains-plugins/ml-experiment-cockpit" rel="noopener noreferrer"&gt;https://github.com/tgg49099-dotcom/atlassian-apps/tree/main/jetbrains-plugins/ml-experiment-cockpit&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If something does not work, write to &lt;code&gt;tgg.apps.support@gmail.com&lt;/code&gt; with your IDE and its build&lt;br&gt;
number (&lt;code&gt;Help | About&lt;/code&gt;), the plugin version, the steps to reproduce the problem, the exact message&lt;br&gt;
shown in the tool window, and the version of your server. Please do not send credentials or&lt;br&gt;
confidential experiment data in a support request.&lt;/p&gt;

&lt;p&gt;One caveat worth stating plainly: JetBrains records your download against the marketplace listing,&lt;br&gt;
and anything you post as a comment or a rating is public on that page. That is JetBrains' own data&lt;br&gt;
flow, not ours.&lt;/p&gt;

&lt;h2&gt;
  
  
  Supporting the maintenance (optional)
&lt;/h2&gt;

&lt;p&gt;The plugin you install is free, and the four tabs above stay free for everyone. What costs me&lt;br&gt;
time is keeping it alive: reading the bug reports, testing against new IDE builds, and keeping up&lt;br&gt;
with the MLflow API changes. If the plugin saved you a few context switches this week and you&lt;br&gt;
want to say thanks, you can leave a small tip here:&lt;br&gt;
&lt;a href="https://tggster.gumroad.com/l/piazge" rel="noopener noreferrer"&gt;https://tggster.gumroad.com/l/piazge&lt;/a&gt; (CHF 3+, whatever you want to give).&lt;/p&gt;

&lt;p&gt;To be precise about what that is: it is &lt;strong&gt;voluntary support, not a purchase&lt;/strong&gt;. It does not unlock&lt;br&gt;
any feature, it does not include support, and the plugin works exactly the same without it. A paid&lt;br&gt;
Pro tier with an extra Compare view is now available on the JetBrains Marketplace (with a&lt;br&gt;
free trial); it is sold there, never through this tip page.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was written with the help of AI and reviewed before publication.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>jetbrains</category>
      <category>mlflow</category>
      <category>python</category>
    </item>
    <item>
      <title>Before you ship: check which model version is live in your MLflow Registry</title>
      <dc:creator>Ti Gg</dc:creator>
      <pubDate>Wed, 11 Nov 2026 16:26:51 +0000</pubDate>
      <link>https://dev.to/ti_gg_043b3e06a4b7d5d0888/before-you-ship-check-which-model-version-is-live-in-your-mlflow-registry-1cd1</link>
      <guid>https://dev.to/ti_gg_043b3e06a4b7d5d0888/before-you-ship-check-which-model-version-is-live-in-your-mlflow-registry-1cd1</guid>
      <description>&lt;p&gt;The moment before a deployment is a bad time to open a browser. You need three facts: which version of the model is currently in Production, whether that version finished registering cleanly, and which training run produced it. The MLflow UI can tell you, but getting there means a tab switch, a login, a tree to expand, and a registry page where you have to hunt for the stage column.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://plugins.jetbrains.com/plugin/35416-ml-experiment-cockpit" rel="noopener noreferrer"&gt;ML Experiment Cockpit&lt;/a&gt; is a free JetBrains plugin that shows the same facts in a tool window next to your code. This tutorial covers one task only: the pre-deployment check, using the free Models tab, plus the same check written as a script you can run in a pipeline.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you need
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;A running MLflow tracking server with a Model Registry, in other words MLflow 2.0 or later. The plugin talks to a server you already run: it never starts one, and it does not read a local &lt;code&gt;mlruns/&lt;/code&gt; directory.&lt;/li&gt;
&lt;li&gt;The plugin, installed from the Marketplace page linked above, in any IntelliJ-platform IDE from build 243 (2024.3) to build 263, which is the range the Marketplace listing states.&lt;/li&gt;
&lt;li&gt;Your tracking URI, for example &lt;code&gt;http://localhost:5000&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Step 1: point the plugin at your server
&lt;/h2&gt;

&lt;p&gt;Open &lt;em&gt;Settings | Tools | ML Experiment Cockpit&lt;/em&gt; and put the URL of your tracking server in the &lt;strong&gt;MLflow tracking URI&lt;/strong&gt; field. This is the same URI you pass to &lt;code&gt;mlflow&lt;/code&gt; commands. The Models tab reuses it, so there is nothing else to configure for this task.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: read the registry
&lt;/h2&gt;

&lt;p&gt;Open &lt;em&gt;View | Tool Windows | ML Experiment Cockpit&lt;/em&gt; and select the &lt;strong&gt;Models&lt;/strong&gt; tab, then click &lt;strong&gt;Refresh&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The left pane lists the registered models your server returns, sorted by name, each one labelled with the number of versions it reports:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;churn-model (2 version(s))
fraud-detector (1 version(s))
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Expand a model. Each version appears with its number and its stage, exactly as the server reports it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;fraud-detector (1 version(s))
  v4 [Production]
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Select the version and the right pane shows the fields that decide whether you ship (the values below are an example):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight yaml"&gt;&lt;code&gt;&lt;span class="na"&gt;name    &lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;fraud-detector&lt;/span&gt;
&lt;span class="na"&gt;version &lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;4&lt;/span&gt;
&lt;span class="na"&gt;stage   &lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Production&lt;/span&gt;
&lt;span class="na"&gt;status  &lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;READY&lt;/span&gt;
&lt;span class="na"&gt;run_id  &lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;8f3c1a9d2b7e4c6f9a0d5e1b2c3d4e5f&lt;/span&gt;
&lt;span class="na"&gt;created &lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;2026-11-10 09:14:22&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;stage&lt;/strong&gt; is the MLflow stage: &lt;code&gt;None&lt;/code&gt;, &lt;code&gt;Staging&lt;/code&gt;, &lt;code&gt;Production&lt;/code&gt; or &lt;code&gt;Archived&lt;/code&gt;. &lt;strong&gt;status&lt;/strong&gt; is the registration status: &lt;code&gt;READY&lt;/code&gt; means the version is registered and ready to be served, while a failed registration would read &lt;code&gt;FAILED_REGISTRATION&lt;/code&gt;. &lt;strong&gt;run_id&lt;/strong&gt; is the training run that produced the version, which is your thread back to the hyperparameters and the metric curves; the Experiments tab of the same tool window shows that run without leaving the IDE.&lt;/p&gt;

&lt;p&gt;If the Models tab answers with an error about the endpoint, your server is older than MLflow 2.0: the Model Registry does not exist before that, and the plugin says so instead of showing you an empty list.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: the same answer, without the IDE
&lt;/h2&gt;

&lt;p&gt;The tab is a convenience over one HTTP GET. This is the request the plugin sends your server when you click &lt;strong&gt;Refresh&lt;/strong&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl -s &lt;span class="s1"&gt;'http://localhost:5000/api/2.0/mlflow/registered-models/search?max_results=500'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A plain GET, with no request body. The answer is JSON with a &lt;code&gt;registered_models&lt;/code&gt; array, and each entry carries a name, a description, a creation timestamp and its &lt;code&gt;latest_versions&lt;/code&gt;. The plugin reads those fields and nothing else. If your server runs with authentication turned on, add your own credentials to the call.&lt;/p&gt;

&lt;p&gt;If you prefer Python, the official MLflow client gives you the same facts, and the same check fits easily into a pipeline:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;mlflow&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;MlflowClient&lt;/span&gt;

&lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;MlflowClient&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"http://localhost:5000"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;search_model_versions&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"name='fraud-detector'"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="k"&gt;print&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current_stage&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;status&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;run_id&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;search_model_versions&lt;/code&gt; takes the filter syntax the registry uses internally, &lt;code&gt;name='&amp;lt;model name&amp;gt;'&lt;/code&gt;, and returns one object per matching version. Each object exposes &lt;code&gt;version&lt;/code&gt;, &lt;code&gt;current_stage&lt;/code&gt;, &lt;code&gt;status&lt;/code&gt;, &lt;code&gt;run_id&lt;/code&gt;, &lt;code&gt;source&lt;/code&gt; and &lt;code&gt;creation_timestamp&lt;/code&gt;, so the one-liner above prints exactly what the Models tab shows.&lt;/p&gt;

&lt;h2&gt;
  
  
  Honest limits
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The Models tab shows the versions your server reports as the latest ones for each model, the &lt;code&gt;latest_versions&lt;/code&gt; field of the response, not the complete history of every version you ever registered. When you need the full list, the script above or the raw GET gives it to you.&lt;/li&gt;
&lt;li&gt;The plugin asks for at most 500 registered models and 500 model versions per query. Those are the plugin's own limits, not MLflow's.&lt;/li&gt;
&lt;li&gt;The Model Registry exists in MLflow 2.0 and later. On an older server, the tab tells you the endpoint is missing instead of showing an empty list.&lt;/li&gt;
&lt;li&gt;The plugin is read-only. It never promotes a version, never tags it, never deletes anything. Promotion stays with the &lt;code&gt;mlflow&lt;/code&gt; client or the MLflow UI.&lt;/li&gt;
&lt;li&gt;Everything in this tutorial is a free feature. No paid tier is involved.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  If it is useful to you
&lt;/h2&gt;

&lt;p&gt;If this plugin saved you a context switch, you can support its development with a free and optional donation: &lt;a href="https://tggster.gumroad.com/l/piazge" rel="noopener noreferrer"&gt;https://tggster.gumroad.com/l/piazge&lt;/a&gt;. It is a donation, not a purchase, and nothing in the free tier depends on it.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was written with the help of AI and reviewed before publication.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>jetbrains</category>
      <category>mlflow</category>
      <category>python</category>
    </item>
    <item>
      <title>ML Experiment Cockpit: read your MLflow, Optuna and W&amp;B runs without leaving your IDE</title>
      <dc:creator>Ti Gg</dc:creator>
      <pubDate>Fri, 30 Oct 2026 07:36:16 +0000</pubDate>
      <link>https://dev.to/ti_gg_043b3e06a4b7d5d0888/ml-experiment-cockpit-read-your-mlflow-optuna-and-wb-runs-without-leaving-your-ide-57aa</link>
      <guid>https://dev.to/ti_gg_043b3e06a4b7d5d0888/ml-experiment-cockpit-read-your-mlflow-optuna-and-wb-runs-without-leaving-your-ide-57aa</guid>
      <description>&lt;p&gt;Reading an ML experiment usually means leaving your editor. You open a browser tab for the MLflow UI, a second one for the W&amp;amp;B dashboard, maybe a notebook cell to call &lt;code&gt;mlflow.search_runs()&lt;/code&gt; and print a metric curve. Each of those is a context switch: a different window, a different login, a different rendering of the same numbers.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://plugins.jetbrains.com/plugin/35416-ml-experiment-cockpit" rel="noopener noreferrer"&gt;ML Experiment Cockpit&lt;/a&gt; is a free, Apache-2.0 JetBrains plugin that puts that information in a tool window inside PyCharm, IntelliJ IDEA, WebStorm, GoLand and CLion. It connects to servers you already run and renders their answers. It never writes anything back.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you get
&lt;/h2&gt;

&lt;p&gt;One tool window, four tabs:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Experiments&lt;/strong&gt; — the MLflow experiment tree: experiments, their runs, each run's status, name and parameters. A &lt;strong&gt;Refresh&lt;/strong&gt; button reloads the tree. Selecting a run opens a detail pane where you pick a metric and read its history as a curve, with the last value and the step it was recorded at.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Optuna&lt;/strong&gt; — studies and trials, read from a running &lt;code&gt;optuna-dashboard&lt;/code&gt;. Multi-objective studies are only partially supported: the first objective is the one displayed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Models&lt;/strong&gt; — the MLflow Model Registry: registered models and their versions, with the stage each version sits in (&lt;code&gt;None&lt;/code&gt;, &lt;code&gt;Staging&lt;/code&gt;, &lt;code&gt;Production&lt;/code&gt;, &lt;code&gt;Archived&lt;/code&gt;). This tab reuses the MLflow tracking URI, so there is nothing extra to configure.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Sweeps&lt;/strong&gt; — Weights &amp;amp; Biases sweeps, listed by state with their run count. The tab asks for a W&amp;amp;B entity and a project name (both fields are pre-filled from the settings) and authenticates with a W&amp;amp;B API key.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Step 1: install it
&lt;/h2&gt;

&lt;p&gt;Search for "ML Experiment Cockpit" in &lt;em&gt;Settings | Plugins&lt;/em&gt; of any JetBrains IDE, or grab it from the &lt;a href="https://plugins.jetbrains.com/plugin/35416-ml-experiment-cockpit" rel="noopener noreferrer"&gt;Marketplace page&lt;/a&gt;. The published versions install on IDEs from 2024.3 (build 243) up to 2026.3 (build 263).&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: point it at your servers
&lt;/h2&gt;

&lt;p&gt;Open &lt;em&gt;Settings | Tools | ML Experiment Cockpit&lt;/em&gt; and fill in what you use:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;MLflow tracking server URL&lt;/strong&gt;, for example &lt;code&gt;http://localhost:5000&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Optuna dashboard URL&lt;/strong&gt;, for example &lt;code&gt;http://localhost:8080&lt;/code&gt;;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;W&amp;amp;B API key&lt;/strong&gt;, plus the default entity and project. The key comes from &lt;a href="https://www.wandb.ai/settings" rel="noopener noreferrer"&gt;https://www.wandb.ai/settings&lt;/a&gt; → Danger Zone → "API keys", and it is stored in the IDE's own credential store, not in a plaintext settings file.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The server-side parts are the ones you already run. The plugin does not start an MLflow server for you.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: read a run
&lt;/h2&gt;

&lt;p&gt;Click &lt;strong&gt;Refresh&lt;/strong&gt; in the Experiments tab and the tree fills with your experiments. Expand one, click a run, and the detail pane lists its parameters and metrics. Pick a metric from the drop-down and the plugin fetches its full history and draws the curve, labelled with the last value and its step.&lt;/p&gt;

&lt;p&gt;Here is the code that does the reading — &lt;code&gt;MlflowClient.getRunMetrics()&lt;/code&gt;, in &lt;code&gt;src/main/kotlin/com/slot5/mlexp/MlflowClient.kt&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight kotlin"&gt;&lt;code&gt;    &lt;span class="k"&gt;fun&lt;/span&gt; &lt;span class="nf"&gt;getRunMetrics&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;runId&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;String&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt; &lt;span class="nc"&gt;Map&lt;/span&gt;&lt;span class="p"&gt;&amp;lt;&lt;/span&gt;&lt;span class="nc"&gt;String&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;Double&lt;/span&gt;&lt;span class="p"&gt;&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="py"&gt;data&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;runGet&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;runId&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="py"&gt;out&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;LinkedHashMap&lt;/span&gt;&lt;span class="p"&gt;&amp;lt;&lt;/span&gt;&lt;span class="nc"&gt;String&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;Double&lt;/span&gt;&lt;span class="p"&gt;&amp;gt;()&lt;/span&gt;
        &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;m&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;data&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;arr&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"metrics"&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="py"&gt;key&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="n"&gt;m&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;str&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"key"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;key&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;isEmpty&lt;/span&gt;&lt;span class="p"&gt;())&lt;/span&gt; &lt;span class="k"&gt;continue&lt;/span&gt;
            &lt;span class="c1"&gt;// MLflow (protobuf JSON) writes NaN / Infinity as *strings*:&lt;/span&gt;
            &lt;span class="c1"&gt;// `m.dbl("value")` would silently turn them into 0.0.&lt;/span&gt;
            &lt;span class="k"&gt;out&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;key&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="p"&gt;=&lt;/span&gt; &lt;span class="n"&gt;m&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;num&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"value"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="o"&gt;?.&lt;/span&gt;&lt;span class="n"&gt;asDouble&lt;/span&gt; &lt;span class="o"&gt;?:&lt;/span&gt; &lt;span class="k"&gt;when&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;m&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;str&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s"&gt;"value"&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
                &lt;span class="s"&gt;"NaN"&lt;/span&gt; &lt;span class="p"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="nc"&gt;Double&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;NaN&lt;/span&gt;
                &lt;span class="s"&gt;"Infinity"&lt;/span&gt; &lt;span class="p"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="nc"&gt;Double&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;POSITIVE_INFINITY&lt;/span&gt;
                &lt;span class="s"&gt;"-Infinity"&lt;/span&gt; &lt;span class="p"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="nc"&gt;Double&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;NEGATIVE_INFINITY&lt;/span&gt;
                &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="p"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="k"&gt;continue&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="k"&gt;out&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Two details are worth reading twice. First, &lt;code&gt;runGet(runId)&lt;/code&gt; is a plain &lt;code&gt;GET /api/2.0/mlflow/runs/get&lt;/code&gt;: MLflow 3.x answers HTTP 405 to a POST on that endpoint, so the verb is measured, not assumed. Second, the NaN branch — MLflow writes &lt;code&gt;NaN&lt;/code&gt; and &lt;code&gt;Infinity&lt;/code&gt; as &lt;em&gt;strings&lt;/em&gt; in its protobuf JSON, and a naive conversion would silently turn a diverged run into a flat 0.0 on your chart. A reader that lies about the shape of your metrics is worse than no reader at all.&lt;/p&gt;

&lt;p&gt;There is no &lt;code&gt;mlflow-client-java&lt;/code&gt; dependency in the plugin. The client is the JDK's &lt;code&gt;HttpClient&lt;/code&gt; plus a small hand-rolled JSON parser of about 6 KB, which avoids pulling in a library with megabytes of transitive dependencies. That parser is deliberately strict: a body that is not JSON raises an error instead of producing an empty list, so "0 experiment(s)" never silently means "the server answered with an HTML error page".&lt;/p&gt;

&lt;h2&gt;
  
  
  Read-only, on purpose
&lt;/h2&gt;

&lt;p&gt;This is a design constraint, not a missing feature. Across the whole plugin there is no &lt;code&gt;PUT&lt;/code&gt;, no &lt;code&gt;DELETE&lt;/code&gt; and no GraphQL mutation. The only requests carrying a body are MLflow's &lt;code&gt;experiments/search&lt;/code&gt; and &lt;code&gt;runs/search&lt;/code&gt;, the "Connect to MLflow Tracking Server" probe (a &lt;code&gt;search&lt;/code&gt; as well), and the W&amp;amp;B GraphQL &lt;code&gt;POST /graphql&lt;/code&gt;, which carries a &lt;code&gt;query&lt;/code&gt; — GraphQL sends queries by POST, and a mutation would be a different document.&lt;/p&gt;

&lt;p&gt;Concretely, the plugin can show you that a run is training, which metric is climbing, whether the model you just registered made it to the Registry, which sweep is running and how many trials it has. It cannot start a run, delete one, tag a model, download an artifact or touch a deployment. If you want to change something, you change it in your own code.&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest limits
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Read-only.&lt;/strong&gt; No writes, no artifacts, no run control, no model serving.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No local MLflow.&lt;/strong&gt; The plugin speaks HTTP to a tracking server you already run (&lt;code&gt;mlflow server&lt;/code&gt;, a container, whatever you use). It does not read the &lt;code&gt;mlruns/&lt;/code&gt; folder of a local run, and it does not ship a server: point the tracking URI at a running instance, otherwise the tabs stay empty. The same goes for Optuna, which needs &lt;code&gt;optuna-dashboard&lt;/code&gt; running as a separate service.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Query limits.&lt;/strong&gt; Up to 500 experiments, 100 runs per experiment, 500 registered models, 500 model versions and 5 000 metric points per run; W&amp;amp;B sweeps are read without pagination, so you get the first 50. These are the limits the plugin asks the server for, not a promise of completeness — with 4 000 runs in an experiment you are looking at the first 100.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you write Python against MLflow every day and would rather read the state of your experiments from the editor you are already in, give it a try and tell us what is missing — the next versions are decided by what you hit.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Disclosure: this article was written with the help of AI and reviewed before publication.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>jetbrains</category>
      <category>mlflow</category>
      <category>python</category>
    </item>
  </channel>
</rss>
