<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Itay Gabbay on Medium]]></title>
        <description><![CDATA[Stories by Itay Gabbay on Medium]]></description>
        <link>https://medium.com/@itaygabbay?source=rss-e08add57fda------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*d8OcaKA4It21jWxrraAGww@2x.jpeg</url>
            <title>Stories by Itay Gabbay on Medium</title>
            <link>https://medium.com/@itaygabbay?source=rss-e08add57fda------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 18:26:57 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@itaygabbay/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Deepchecks ❤️ Weights & Biases for Testing ML within the Training Workflow]]></title>
            <link>https://medium.com/@itaygabbay/deepchecks-%EF%B8%8F-weights-biases-for-testing-ml-within-the-training-workflow-9f509f7dfbcc?source=rss-e08add57fda------2</link>
            <guid isPermaLink="false">https://medium.com/p/9f509f7dfbcc</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ml-testing]]></category>
            <category><![CDATA[deepchecks]]></category>
            <category><![CDATA[model-validation]]></category>
            <dc:creator><![CDATA[Itay Gabbay]]></dc:creator>
            <pubDate>Thu, 10 Mar 2022 12:09:36 GMT</pubDate>
            <atom:updated>2022-03-10T13:50:08.191Z</atom:updated>
            <content:encoded><![CDATA[<h4>Deepchecks and W&amp;B have collaborated to create an integrated experience of experiment tracking and ML validation. This is done by running deepchecks’ test suites and logging the results to W&amp;B. Here’s how:</h4><p>ML Model validation, especially in production environments, has always been one of the most difficult subtopics in machine learning research. Writing tests for ML models is a challenging task and must take into account a lot of aspects, both in the data and in the model itself, which might get really exhausting.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*u5ka36fzZ6RD7jV9R1OBmA.gif" /><figcaption>Example of a test suite output</figcaption></figure><p>Deepchecks (<a href="https://github.com/deepchecks/deepchecks">https://github.com/deepchecks/deepchecks</a> ⭐️)is an open-source Python package for comprehensively validating your machine learning models and data with minimal effort. This includes checks related to various types of issues, such as model performance, data integrity, distribution mismatches, and more.</p><p>In the latest release (0.5.0), deepchecks support tracking the suite’s results with <a href="https://wandb.ai/site">Weights &amp; Biases</a>. To give you some context, Weights &amp; Biases helps you keep track of your machine learning experiments. You can use the tool to log output metrics from your runs, and then visualize and compare results and quickly share findings with your colleagues.</p><p>In this tutorial, we will through an example of sending a deepchecks suite’s output to W&amp;B.</p><p>First, let’s install the deepchecks package.</p><pre>pip install deepchecks</pre><p>After the installation is done, we will prepare the train &amp; test sets along with a model. In this tutorial we will use the same <a href="https://docs.deepchecks.com/en/stable/examples/guides/quickstart_in_5_minutes.html">example from deepchecks’ docs</a>, using the Iris dataset:</p><pre><em># General imports</em><br><strong>import</strong> pandas <strong>as</strong> pd<br><strong>import</strong> numpy <strong>as</strong> np<br><br><strong>from</strong> deepchecks.datasets.classification <strong>import</strong> iris<br><strong>from</strong> sklearn.ensemble <strong>import</strong> RandomForestClassifier<br><strong>from</strong> sklearn.model_selection <strong>import</strong> train_test_split<br><br><em># Load Data</em><br>iris_df <strong>=</strong> iris<strong>.</strong>load_data<strong>(</strong>data_format<strong>=</strong>&#39;Dataframe&#39;<strong>,</strong> as_train_test<strong>=False)</strong><br>label_col <strong>=</strong> &#39;target&#39;<br>df_train<strong>,</strong> df_test <strong>=</strong> train_test_split<strong>(</strong>iris_df<strong>,</strong> stratify<strong>=</strong>iris_df<strong>[</strong>label_col<strong>],</strong> random_state<strong>=0)</strong><br><br><em># Train Model</em><br>rf_clf <strong>=</strong> RandomForestClassifier<strong>()</strong><br>rf_clf<strong>.</strong>fit<strong>(</strong>df_train<strong>.</strong>drop<strong>(</strong>label_col<strong>,</strong> axis<strong>=1),</strong> df_train<strong>[</strong>label_col<strong>])</strong></pre><pre><strong>from</strong> deepchecks <strong>import</strong> Dataset<br><br><em># We explicitly state that this dataset has no categorical features, otherwise they will be automatically inferred</em><br><em># If the dataset has categorical features, the best practice is to pass a list with their names</em><br><br>ds_train <strong>=</strong> Dataset<strong>(</strong>df_train<strong>,</strong> label<strong>=</strong>label_col<strong>,</strong> cat_features<strong>=[])</strong><br>ds_test <strong>=</strong>  Dataset<strong>(</strong>df_test<strong>,</strong>  label<strong>=</strong>label_col<strong>,</strong> cat_features<strong>=[])</strong></pre><p>After we’ve prepared the datasets, we will import the full suite and run it:</p><pre><strong>from</strong> deepchecks.suites <strong>import</strong> full_suite<br><br>suite <strong>=</strong> full_suite<strong>()<br></strong>suite_result = suite.run(ds_train, ds_test, rf_clf)</pre><p>Now, all results from the test suite are calculated and the last step will be to send those to Weight&amp;Biases:</p><pre>suite_result<strong>.</strong>to_wandb<strong>()</strong></pre><p><strong>And that’s it!</strong></p><p>Let’s explore a bit how the results are tracked inside W&amp;B:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*G_b_jg8722CW6dZNTa8vnA.gif" /><figcaption>Suite output in W&amp;B</figcaption></figure><p>We can see that each check is contained within a section in the W&amp;B dashboard, which enables us to compare between different runs. This way, we can easily compare between models and select the best one</p><p><strong>Next steps</strong></p><p>You should attempt to fork <a href="https://docs.deepchecks.com/en/stable/examples/guides/export_outputs_to_wandb.html">this notebook</a> to learn more, integrate it with your dataset, and try to compare different results with W&amp;B. Good luck!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9f509f7dfbcc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Feature Toggles — or why you should take risks in production]]></title>
            <link>https://medium.com/@itaygabbay/feature-toggles-or-why-you-should-take-risks-in-production-cad9f22565f1?source=rss-e08add57fda------2</link>
            <guid isPermaLink="false">https://medium.com/p/cad9f22565f1</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[feature-toggles]]></category>
            <category><![CDATA[continuous-delivery]]></category>
            <category><![CDATA[agile]]></category>
            <dc:creator><![CDATA[Itay Gabbay]]></dc:creator>
            <pubDate>Wed, 21 Feb 2018 20:53:05 GMT</pubDate>
            <atom:updated>2018-02-21T20:53:05.224Z</atom:updated>
            <content:encoded><![CDATA[<p>In today’s software engineering world, where continuous integration and deployment is not just a buzzword anymore, where every line of code goes down to production in seconds, and sometimes without a manual check of a person, any change in your codebase becomes dangerous and may put your application at risk. Sometimes, this CI/CD approach may also introduce a feature that is not fully ready or even new bugs to your PRODUCTION environment.</p><p>The fear of “losing control” on the production environment make many large organizations and managers avoid implementing the continues deployment approach in their applications and products.</p><p>If only you could manage all the new features that come in every day and decide with the click of a button whether to expose them to users or not, or control which lines of code will run, we would all be more relaxed and in control of our systems….</p><p>This dreamy idea is possible, and it is even quite simple just like switching a toggle.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/1*kt0SgPdDlP_g0qvQqwYc3A.jpeg" /></figure><p>“Feature Toggle” is a software development approach that allows development teams to meet demands and deliver new features to the users quickly but with <strong>no risk</strong>. It is made possible by managing flags that are inserted into the new feature code and allowing changes in the way the code runs in the production environment without deploying an updated version.</p><p>Let’s illustrate this with a simple example:</p><p>Let’s say you are an engineer on one of Facebook’s development teams. There’s a new requirement: you need to improve the Like button for posts, and you know it’s a long development process that will take a few weeks. During this time your team needs to work on parts of code related to the tremendous change you are going to make. To avoid a complicated big branch, because you have past painful experience of code merge that takes a few days, you have decided that the team will continue their normal tasks, and engineers working on the Like button will use <strong>feature toggle,</strong> so they don’t ruin or delay the rest of the team work and don’t damage the quality of the release. Here’s how the code looks:</p><p>before:</p><p>function likePost(){</p><p>// current implementation</p><p>}</p><p>after:</p><pre>function likePost(){</pre><pre>  var useNewAlgorithm = false;</pre><pre>  // useNewAlgorithm = true; // UNCOMMENT FOR NEW LIKE ALGORITHM</pre><pre>  if( useNewAlgorithm ){</pre><pre>    return newLikePost();</pre><pre>  } else{</pre><pre>    return oldLikePost();</pre><pre>  }</pre><pre>}</pre><pre>function oldLikePost (){</pre><pre>  // current implementation lives here</pre><pre>}</pre><pre>function newLikePost (){</pre><pre>  // TODO: implement better Like button algorithm</pre><p>The toggle can be improved in several ways and even shortened to one line of code:</p><pre>function likePost(){</pre><pre>  return isFeatureEnabled(‘NewLikeAlgorithm’, newLikePost(),     oldLikePost());</pre><pre>}</pre><p>Note the isFeatureEnabled operation, which is used as a “Toggle router” and can be used in every code section and on any toggle in the application in order to control which code to run.</p><p>Now that you’ve developed the new Like button, you’ve decided to expose it to users. Here too there are several approaches and methods. One option is to enable Toggle for all users of the system, and everyone will get the new feature. If you find a problem with the feature, or a bug, you can turn the toggle off and the old logic will come back. Just like turning off the light in your office. A second option is to expose the Toggle to specific users / groups, so that the new feature will be exposed to them only, and the rest of the users will continue to use the old feature. This allows you to get feedback on the new feature and decide whether to expose it to everyone.</p><p>Once we understand what Feature Toggle is and why you should use it, the question is how do you implement it in your applications?</p><p>There are several options to implement the Feature toggle:</p><p>• There are several tools that manage all the toggles of your app, and you can turn them on and off through an interface the tool expose. All you need to do is combine the toggle with the code. But — most of them are not free and will cost you some money.</p><p>• You can develop a feature toggle management tools yourself and embed it in your application. It’s a great challenge to your senior engineers 😊</p><p>An example of the Feature toggle management interface:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fLOMneA2sWiCpLcf7TrfoA.jpeg" /><figcaption>LaunchDarkly toggle’s interface</figcaption></figure><p>In my organization, we have developed an internal feature toggle management app that manages toggles to all users and to specific users &amp; groups, and even to a specific user. It was developed as an external service so that any application in our organization can use it and manage its flags.</p><p>It is important to note that after a while (depending on the size of your application), managing the toogles become more and more difficult because their number just go up. You can compare it to a warehouse that just messes up with every object you put in it until you cannot find what you were looking for. To avoid this, it is recommended to keep this warehouse the most organized ever, i.e. once you no longer need a toggle, the best thing to do is to delete it.</p><p>And in conclusion — illustration of everything we’ve talked about here:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PS4bAZhd1f5Daf0rGhVdMw.png" /><figcaption>Spotify feature toggle approach</figcaption></figure><p>Good luck implementing your toggles!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cad9f22565f1" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>