<?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 Kasun Bandara on Medium]]></title>
        <description><![CDATA[Stories by Kasun Bandara on Medium]]></description>
        <link>https://medium.com/@kasun-bandara?source=rss-6b3c2920e29------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*nt5wUId4lhThyNjO31HvRQ.jpeg</url>
            <title>Stories by Kasun Bandara on Medium</title>
            <link>https://medium.com/@kasun-bandara?source=rss-6b3c2920e29------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 20:45:08 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kasun-bandara/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[Rewrite URI using a regex pattern]]></title>
            <link>https://kasun-bandara.medium.com/rewrite-uri-using-a-regex-pattern-93a7d3c3cbfc?source=rss-6b3c2920e29------2</link>
            <guid isPermaLink="false">https://medium.com/p/93a7d3c3cbfc</guid>
            <category><![CDATA[nginx]]></category>
            <category><![CDATA[url-rewriting]]></category>
            <category><![CDATA[regex]]></category>
            <category><![CDATA[nginx-ingress]]></category>
            <dc:creator><![CDATA[Kasun Bandara]]></dc:creator>
            <pubDate>Mon, 01 Mar 2021 23:07:10 GMT</pubDate>
            <atom:updated>2021-03-01T23:30:15.136Z</atom:updated>
            <content:encoded><![CDATA[<h4>Nginx Ingress</h4><h4>Evaluate different approaches to rewrite the URI based on the subdomain of Nginx Ingress</h4><p>Sometimes you may need to use a wild card domain name in the Nginx ingress and route traffic by rewriting the URI based on the subdomain name. For example, consider wildcard domain <strong>*.domain.com</strong>;</p><pre>foo-bar.domain.com &gt; service-name:port/foo/bar</pre><p>If we use the default Nginx server, this is pretty much straight forward and you can simply use the following code block in order to rewrite the URI based on the subdomain.</p><pre>server {     <br>    server_name ~^(?&lt;prefix&gt;\w+)-(?&lt;suffix&gt;\w+)\.domain\.com$;</pre><pre>    location / {         <br>        proxy_pass http://service_upstream/$prefix/$suffix$uri;     <br>    } <br>}</pre><p>But when it comes to the Nginx ingress it’s not the case. It’s difficult to generate the above configuration using the Ingress resource due to the limitation in using regex syntax for the hostname.</p><blockquote>“Kubernetes only accept expressions that comply with the RE2 engine syntax. It is possible that valid expressions accepted by NGINX cannot be used with ingress-nginx, because the PCRE library (used in NGINX) supports a wider syntax than RE2.” [1]</blockquote><p>Before we move forward, let&#39;s clear out our understanding of how the Nginx configuration is generated by the Kubernetes Ingress controller.</p><p>Nginx configuration consists of two main block directives called <em>event</em> and <em>http</em>, where <em>server</em> and <em>location</em> block directives reside in the <em>http</em> main block. Server block defines a virtual server and used to handle the request based on hostname, IP address and port number. Location block resides withing the server block and used to define the routing based on the URI of the parent server.</p><p>Single Nginx ingress resource represents a server block in Nginx configuration and supports adding multiple <em>paths</em> which is equivalent to the <em>location</em> block in the Nginx configuration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/541/1*trmEVe-ZX949dahwKLJOPA.png" /><figcaption>How Nginx configuration generated from Nginx ingress resource.</figcaption></figure><p>Now, let&#39;s find out several approaches on how we can create a regex hostname comply with <em>PCRE</em> syntax.</p><h4><strong>Adding regex pattern to the default server block</strong></h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/223d3d37fa97de749241eef7f3978365/href">https://medium.com/media/223d3d37fa97de749241eef7f3978365/href</a></iframe><p>Here we have added regex pattern using server-snippet[2] and the rewrite logic using rewrite-target[3] annotations. Since we have not defined a hostname in the rule defined(line 15), this configuration will be added to the default server block. Note that you cannot override the hostname by adding a server snippet if you define a hostname in the ingress rules. So it has to be the default server block if you need to define a hostname with the regex expression with <em>PCRE</em> syntax.</p><p><strong>Drawbacks:<br></strong>As you can see here, we have overridden the default server block so all the traffic which doesn’t comply with any Nginx rules end up here. And this is not recommended because the default server block is there to serve a different purpose.</p><blockquote>Nginx tests only the request’s header field “Host” to determine which server the request should be routed to. If its value does not match any server name, or the request does not contain this header field at all, then nginx will route the request to the default server block.</blockquote><p>You may get into trouble if you have pointed out multiple domains’ DNS records to the same Nginx server. In case if you forget to define all rules related to those domains, some requests will be handled by the default server block which is now overridden to serve other wildcard domain name.</p><h4><strong>Adding URI rewrite logic using a Lua Script</strong></h4><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/c0e9d48cd9947dbb9962cbf9e53ce0a3/href">https://medium.com/media/c0e9d48cd9947dbb9962cbf9e53ce0a3/href</a></iframe><p>Nginx ingress by default uses <em>lua-nginx-module</em>[4] to avoid reloading Nginx in every upstream change. Otherwise, Nginx reloads will affect response latency and the quality of the load balancing(Nginx reset the load balancing state in every reload).</p><p>Because of this, we can execute the <em>Lua</em> code blocks inside the Nginx configuration out of the box using the <em>configuration-snippet</em>[5] annotation. In this case, we’ll use <em>set_by_lua_block</em> module which is used to execute short and fast running code blocks inside Nginx. But most importantly, you need to make sure the efficiency of the code block running since the Nginx event loop is blocked during the code execution.</p><p>Drawbacks:<br>The major drawback of this approach is that complex Lua logic will slow down the overall performance of the Nginx controller. <br>Due to restricted permissions inside the <em>set_by_lua_block</em>, error handling is not possible. So every error will trigger a 503 response code from the Nginx.</p><h4><strong>Adding URI rewrite logic using a Map directive</strong></h4><p>This is the recommended approach considering the drawbacks of the aforementioned approaches. A <em>map</em> directive[6] is an Nginx way of defining a method; a <em>map</em> creates a new variable whose value depends on the values of one or more of the source variables specified in the first parameter.</p><pre>map $http_host $prefix { <br>  ~(?&lt;prefix&gt;\\w+)-(?&lt;suffix&gt;\\w+)\\.domain\\.com  $prefix; <br>  default  ‘’; <br>} </pre><p>In the above code block, we have defined a <em>map</em> directive that will use the <em>$http_host</em> as an input variable(a default variable used in Nginx) and output the <em>$prefix</em> and <em>$suffix</em> as the output variables. And we can use the defined <em>map</em> directive inside the ingress definition as depicted in the below code block (line 7).</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/15a1defde8a4805678f7fbf2910c2e5b/href">https://medium.com/media/15a1defde8a4805678f7fbf2910c2e5b/href</a></iframe><p>Since the <em>map</em> directive should be defined under <em>HTTP</em> context of the Nginx, we cannot directly inject them into the Nginx ingress definition. Nginx ingress only allows you to create and customize the <strong><em>location</em></strong> and<strong><em> server</em></strong> directives.</p><p>However, the Nginx controller allows you to customize the Nginx configuration using a <em>ConfigMap</em> resource. So in this case we can use <em>http-snippet</em>[7] config in order to add additional configuration to the <em>HTTP</em> context of the Nginx configuration. So we can add our <em>map</em> directive using <em>http-snippet </em>to the Nginx <em>configMap</em> as shown below,</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/e6f64980b40ff53abdba18621ab637a7/href">https://medium.com/media/e6f64980b40ff53abdba18621ab637a7/href</a></iframe><p>A huge advantage of using this approach is that variables defined using <em>map</em> directives will only be executed when they are used. So even a large number of <em>map</em> directive does not add performance overhead to the request processing.</p><p>So my final thought on this is that always consider the performance and security implications before you choose any approach mentioned above.</p><p><strong>Resources:<br></strong>[1] <a href="https://kubernetes.github.io/ingress-nginx/user-guide/ingress-path-matching/">https://kubernetes.github.io/ingress-nginx/user-guide/ingress-path-matching/</a><br>[2] <a href="https://kubernetes.github.io/ingress-nginx/user-guide/nginx-configuration/annotations/#server-snippet">https://kubernetes.github.io/ingress-nginx/user-guide/nginx-configuration/annotations/#server-snippet</a><br>[3] <a href="https://kubernetes.github.io/ingress-nginx/user-guide/nginx-configuration/annotations/#rewrite">https://kubernetes.github.io/ingress-nginx/user-guide/nginx-configuration/annotations/#rewrite</a><br>[4] <a href="https://github.com/openresty/lua-nginx-module">https://github.com/openresty/lua-nginx-module</a><br>[5] <a href="https://kubernetes.github.io/ingress-nginx/user-guide/nginx-configuration/annotations/#configuration-snippet">https://kubernetes.github.io/ingress-nginx/user-guide/nginx-configuration/annotations/#configuration-snippet</a><br>[6] <a href="http://nginx.org/en/docs/http/ngx_http_map_module.html#map">http://nginx.org/en/docs/http/ngx_http_map_module.html#map</a><br>[7] <a href="https://docs.nginx.com/nginx-ingress-controller/configuration/global-configuration/configmap-resource/#snippets-and-custom-templates">https://docs.nginx.com/nginx-ingress-controller/configuration/global-configuration/configmap-resource/#snippets-and-custom-templates</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=93a7d3c3cbfc" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Demystifying Linux 101: File System and the permission conflict.]]></title>
            <link>https://kasun-bandara.medium.com/demystifying-linux-101-file-system-and-the-permission-conflict-e04588843ff2?source=rss-6b3c2920e29------2</link>
            <guid isPermaLink="false">https://medium.com/p/e04588843ff2</guid>
            <category><![CDATA[file-permission]]></category>
            <category><![CDATA[inode]]></category>
            <category><![CDATA[terminal]]></category>
            <category><![CDATA[filesystem]]></category>
            <dc:creator><![CDATA[Kasun Bandara]]></dc:creator>
            <pubDate>Mon, 31 Dec 2018 08:12:19 GMT</pubDate>
            <atom:updated>2020-03-15T13:36:43.107Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Demystifying Linux 101: File System</strong></h3><p><strong>“Everything is a file”</strong></p><p>This is one of the key concepts that drove the Unix operating system to the success. A file provides a common abstraction to the wide range of I/O resources such as documents, directories, hard-drives, CD-Roms, modems, keyboards, printers, monitors, terminals and even some inter-process and network communications.</p><p>•Your computer memory is a file object located in /dev/mem.<br>•Your first hard disk is a file object located in /dev/sda.<br>(Note that Linux kernel is responsible for exposing those information to use as a file.)</p><h4>Inode</h4><p>The <strong>inode</strong> (index node) is a data structure in a Unix-style file system that describes a file-system object such as a file or a directory. Each inode stores the attributes and disk block locations of the object’s data. An inode contains a <strong>File Controle Block (FCB) </strong>which has the details about permissions, owner, group, size, access/modify times, etc related to the content stored inside the inode. Inode content is stored in the disk blocks and the inode has pointers that refer to this disk blocks.</p><p>FCB contains the following information.</p><ul><li><strong>File type:</strong> a regular file or a directory</li><li><strong>Permissions to that file:</strong> read, write, execute</li><li><strong>Link count: </strong>The number of hard links relative to an inode</li><li><strong>User ID:</strong> the owner of the file</li><li><strong>Group ID:</strong> group owner</li><li><strong>Size of file:</strong> or major/minor number in case of some special files</li><li><strong>Timestamp</strong>: access time, modification time and (inode) change time</li><li><strong>Attributes:</strong> immutable’ for example</li><li><strong>Access control list:</strong> permissions for special users/groups</li><li><strong>Link to the location of the file stored in a disk block</strong></li><li><strong>Other metadata </strong>about the file</li></ul><p>As you have noticed in the above FCB does not contain a name attribute. So that we can have multiple names for a single inode and that leads to the concept called hard-link which will discuss later.</p><p>You may have noticed that there is a property called “File type” in the FCB. This is to denote whether the inode is a file or a directory object.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/665/1*hpgIbNN4cdn938qLva5i7Q.png" /></figure><p>If we consider tree structure of the Linux file system, we can denote it as the above diagram. Where root itself is an inode directory that stores the all other file and directory information. “file1” is a file inode located in the /home/user/ directory path. When a file1 is created, it is assigned both <strong>name</strong> and an <strong>inode number</strong> which is <strong>unique to the file system</strong>. Both name and the inode number is stored inside the directory (user inode). A directory inode holds only the<strong> names of the files that appear to the user </strong>and <strong>their inode numbers</strong>, the file inode(file1) contains all the other information describing a file. And the name that references an inode is called hard-link. So the main thing here need to be clarified is that a file’s name belongs to the directory inode and the file content is stored in a separate file inode.</p><p>If you want to find out attribute information of a file, that information is stored in the inode of the object (file inode), not in the directory inode. You must first use the inode number associated with the file object to find the file attributes.</p><p><strong>“ls” or “ls -i” </strong>command<strong> </strong>only need to read the names and inode numbers from the directory — no additional inode access is needed because no other attributes are being queried. Reading the one directory inode is sufficient.</p><p><strong>“ls -l” </strong>command has to display attribute information for every object named in the directory, so it has to do a separate inode lookup to find out the inode attribute information for every inode in the directory. A directory with 1000 names in it requires 1000 separate inode lookups to fetch the attributes!</p><p>(This is why ls or ls -i are much faster than ls -l on a large directory)</p><p>We can use the “stat” command to retrieve the inode information of a file as shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/719/1*QlRTPIA6PJpA4TQPONHzgA.png" /></figure><h4>File permission conflicts associated with this behavior.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/607/1*7kGNlZYWNtCakxORnnQYVQ.png" /></figure><p>In this example, I have created a directory “dir1” which has the read and execute permission for the current user. (d r-x rwx r-x denotes that its a directory, the<strong> user has r-x permission</strong>, the <strong>user group has rwx permission </strong>and<strong> others have r-x permission.</strong>) And a text file “file1.txt” which has read-write permission to the current user “kasun” inside the “dir1”.</p><p>But when I try to delete the “file1.txt” inside “dir1” directory, permission was denied even though the “file1.txt” has the write permission to delete.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*rezMJGO_Q4SygyMXNlhJCA.png" /></figure><p>The reason for this behavior is that the “dir1” directory inode which stores the file name does not have the write permission to delete the <strong>file name.</strong></p><p>Let&#39;s add write permission to the directory and remove all the permission in the file for the current user.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/712/1*h-bk0Wg7QM3drT5MaVnbWQ.png" /></figure><p>Then try to delete the “file1.txt”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/662/1*3ywe7raXXQ4zoO61sPjh_g.png" /></figure><p>Here we were able to delete the “file1.txt” from the “dir1” directory by giving write permission to the directory which holds the file even though the user has no permission to the file.</p><p>The link between the file name and the inode is called a <strong>hard-link</strong>. Since file name and the inode are separate entities, one inode can have multiple names. The number of links that an inode has denotes by the <strong>link-count </strong>in FCB. The ln command can create a new name (a new hard link) in a directory for an existing file inode, increasing the file’s inode link count. The rm command removes a name (a hard link) from a directory, decreasing the file’s inode link count.</p><p>When the link count for an inode goes to zero, the inode has no names and the inode is freed and recycled and all the storage and data used by the item are released.</p><p>The rm command does not remove files; it removes names for files. When all the names for a file inode are removed, the system removes the inode itself and releases all the disk space.</p><p>As long as an inode has at least one name in some directory (a non-zero link count), it cannot be freed up and released.</p><p>References;</p><p><a href="http://teaching.idallen.com/cst8207/13w/notes/450_file_system.html">Unix/Linux File System - directories, inodes, hard links</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e04588843ff2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Streaming ML: Perceptron Algorithm]]></title>
            <link>https://kasun-bandara.medium.com/steaming-ml-perceptron-algorithm-67bf29771ed9?source=rss-6b3c2920e29------2</link>
            <guid isPermaLink="false">https://medium.com/p/67bf29771ed9</guid>
            <category><![CDATA[perceptron]]></category>
            <category><![CDATA[streaming-analytics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Kasun Bandara]]></dc:creator>
            <pubDate>Wed, 05 Apr 2017 19:06:15 GMT</pubDate>
            <atom:updated>2017-04-16T04:20:41.329Z</atom:updated>
            <content:encoded><![CDATA[<p>Machine learning has become mainstream in big data analysis and improving exponentially breaking lot of barriers with the help of modern computation capability. As the vast popularity of Internet of Things, large amount of data is acquired over time. Instead of waiting until the data is collected, streaming data analysis bring us a great opportunity to analyse and predict the data on the spot. When the data change over the time streaming analysis adapt and scale when it’s impractical to store raw data and while persist on smaller and focused representation.</p><p><strong>Batch Processing vs Streaming</strong></p><p>Batch processing is used for the large volume of data while streaming is suitable for data with high velocity. As the growth of the IoT, large amount of sensor data is gathered and shared rapidly overtime. So there is a huge demand has emerged for the streaming data analysis. Batch processing is efficient if we used method like mapreduce to distribute the data across the clusters to be processed in batch processing. But when it comes to streaming data, it should be able to handle data with extremely <strong>low latency</strong> for workloads that must be processed with <strong>minimal delay</strong>.</p><p>Traditional Data Mining Algorithms work with a static dataset and the algorithm can afford to read the data several times on the other hand Stream Mining only can afford to read the data once so the algorithms for this subfield of data mining are based on a single scan. Examples of data streams include computer network traffic, phone conversations, ATM transactions, web searches, and sensor data.</p><h3><strong>Implement the Perceptron</strong></h3><p><strong>Building the hypothesis function h(x)</strong></p><p>x(i) = input variable<br>y(i) = target variable</p><p>approximate y as a linear function of x:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/262/0*84zqWtG6JzwhaFPY." /></figure><p>h(x) = predicted value<br>θ = Weights values</p><p>let x0 = 1 (intercept term)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/247/0*5FVqWYBKvqF8fUWI." /></figure><p><strong>Cost Function</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/267/0*xumcTRxHPpcpjfiX." /></figure><p>Here we use gradient descent algorithms to minimize the cost function.</p><p><strong>gradient descent algorithm</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/201/0*r3ZgYKiIELTRMJ3r." /></figure><p>α = Learning Rate</p><p>But as the Widrow-Hoff learning rule or LMS update rule for a single training example:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/334/1*tG9tJPUgvhKfG1biEiL0Sg.png" /></figure><p>Therefore</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/0*wnhPFQCXotp6orR3." /></figure><p><strong>Stochastic Gradient Descent (incremental gradient descent)</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/483/0*gmqjXV9sx9NISVRx." /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/211/1*PU0BzNRb1SUpeb8i6dpy5Q.png" /></figure><p>In this algorithm, we repeatedly run through the training set, and each time we encounter a training example, we update the parameters according to the gradient of the error with respect to that single training example only. This algorithm is called stochastic gradient descent (also incremental gradient descent). Whereas batch gradient descent has to scan through the entire training set before taking a single step — a costly operation if m is large — stochastic gradient descent can start making progress right away, and continues to make progress with each example it looks at. Often, stochastic gradient descent gets θ “close” to the minimum much faster than batch gradient descent. (Note however that it may never “converge” to the minimum, and the parameters θ will keep oscillating around the minimum of J(θ); but in practice most of the values near the minimum will be reasonably good approximations to the true minimum.) For these reasons, particularly when the training set is large, stochastic gradient descent is often preferred over batch gradient descent.</p><p>However, it has been shown that SGD almost surely converges to the global cost minimum if the cost function is convex (or pseudo-convex).</p><p><strong>Improve Stochastic Gradient Descent based learning</strong></p><ul><li>An adaptive learning rate η Choosing a decrease constant d that shrinks the learning rate over time:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/190/0*NZuRCuzYADXPLY3F." /></figure><ul><li>Momentum learning by adding a factor of the previous gradient to the weight update for faster updates:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/212/0*ISRNTerO0vpF5Vc6." /></figure><ul><li>Randomly shuffle samples in the training set.</li></ul><p><strong>Coding…</strong></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/f2317ae67f3e7b84156b6f1191612571/href">https://medium.com/media/f2317ae67f3e7b84156b6f1191612571/href</a></iframe><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/1fbe6e30f2b8d87a0bb7a5378aadda5a/href">https://medium.com/media/1fbe6e30f2b8d87a0bb7a5378aadda5a/href</a></iframe><p>I used Iris Data Set from UC Irvine Machine Learning Repository for the calculate the accuracy of the model. Which can be downloaded from this <a href="https://archive.ics.uci.edu/ml/datasets/Iris">link</a></p><p>Note — As this dataset is used for the binary classification, I preprocessed the dataset accordingly.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=67bf29771ed9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Build a OS resume]]></title>
            <link>https://kasun-bandara.medium.com/build-a-os-resume-d6b15ef0378d?source=rss-6b3c2920e29------2</link>
            <guid isPermaLink="false">https://medium.com/p/d6b15ef0378d</guid>
            <dc:creator><![CDATA[Kasun Bandara]]></dc:creator>
            <pubDate>Sun, 26 Feb 2017 06:58:59 GMT</pubDate>
            <atom:updated>2017-02-26T15:21:23.393Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UgJOdbVDc77JFhuQg4gYww.png" /></figure><p>To be continued…</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d6b15ef0378d" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>