<?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 Márton Kodok on Medium]]></title>
        <description><![CDATA[Stories by Márton Kodok on Medium]]></description>
        <link>https://medium.com/@martonkodok?source=rss-350650c446a8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*IxybV6BwVbbruD_9xweD1Q.jpeg</url>
            <title>Stories by Márton Kodok on Medium</title>
            <link>https://medium.com/@martonkodok?source=rss-350650c446a8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 13:02:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@martonkodok/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[PHP on Cloud Run: Cut Your Latency with PHP 8.5 and Persistent Sockets]]></title>
            <link>https://medium.com/google-cloud/php-on-cloud-run-cut-your-latency-with-php-8-5-and-persistent-sockets-7ff749fd02a9?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ff749fd02a9</guid>
            <category><![CDATA[php]]></category>
            <category><![CDATA[google-cloud-run]]></category>
            <category><![CDATA[curl]]></category>
            <category><![CDATA[gcp-app-dev]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 10:45:25 GMT</pubDate>
            <atom:updated>2026-10-19T05:34:00.851Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SagMpEHCQhrql5zjsPxAHw.png" /></figure><p>When <strong>PHP </strong><strong>8.5.0</strong> shipped with the new curl_share_init_persistent() engine primitive, I opened <a href="https://github.com/guzzle/guzzle/issues/3307"><strong>guzzle/guzzle#3307</strong></a> on GitHub to propose modifying Guzzle&#39;s cURL transport layer so PHP applications could natively pool and reuse live sockets across requests. Now that the Guzzle team has officially shipped native persistent transport sharing in <strong>Guzzle </strong>(and the <strong>Google Cloud PHP SDK</strong> has updated its HTTP handler factory to support Guzzle 8), I was excited to finally take it for a spin on a real multi-service Cloud Run workload.</p><p>Even when a Cloud Run instance is warm and serving dozens of requests per second, standard PHP request lifecycles tear down their socket instances at the end of every HTTP execution—destroying the underlying TCP sockets, TLS 1.3 state, and HTTP/2 streams to *.googleapis.com.</p><p>Before diving into the architecture, look at the raw HTTP response headers from two consecutive requests hitting our API endpoint (which touches <strong>Google OAuth2</strong>, <strong>BigQuery</strong>, and <strong>Cloud Datastore</strong> in a single call):</p><p>Cold Worker — Opens Sockets &amp; Seeds CurlSharePersistentHandle</p><pre>X-Transport-Trace-1: host=oauth2.googleapis.com local_port=51412 dns=34.33ms tcp=27.03ms tls=92.22ms ttfb=202.62ms total=203.27ms socket_reused=NO (Cold Handshake)<br>X-Transport-Trace-2: host=bigquery.googleapis.com local_port=51415 dns=32.43ms tcp=57.01ms tls=175.31ms ttfb=1653.11ms total=1667.24ms socket_reused=NO (Cold Handshake)<br>X-Transport-Trace-3: host=datastore.googleapis.com local_port=51429 dns=72.66ms tcp=50.61ms tls=153.86ms ttfb=1409.00ms total=1409.23ms socket_reused=NO (Cold Handshake)</pre><p>Warm Worker — Reuses All 3 Sockets with 0.00ms Handshakes</p><pre>X-Transport-Trace-1: host=oauth2.googleapis.com local_port=51412 dns=0.07ms tcp=0.00ms tls=0.00ms ttfb=55.57ms total=56.05ms socket_reused=YES (0ms Handshake)<br>X-Transport-Trace-2: host=bigquery.googleapis.com local_port=51415 dns=0.02ms tcp=0.00ms tls=0.00ms ttfb=632.90ms total=652.75ms socket_reused=YES (0ms Handshake)<br>X-Transport-Trace-3: host=datastore.googleapis.com local_port=51429 dns=0.03ms tcp=0.00ms tls=0.00ms ttfb=417.27ms total=417.50ms socket_reused=YES (0ms Handshake)</pre><p>Every single local port (51412, 51415, 51429) survived the end of Request #1 and was reused by Request #2 with <strong>tcp=0.00ms and </strong><strong>tls=0.00ms</strong>, dropping combined Google Cloud transport time from <strong>3,279.74 ms down to </strong><strong>1,126.30 ms</strong>. Here is how the feature evolved, the hidden libcurl gotchas we uncovered in Guzzle 8.2&#39;s source code, and how to wire it into your Cloud Run stack.</p><h3>The Evolution: How Persistent Transport Sharing Landed</h3><p>Solving cross-request connection pooling in PHP required alignment across three layers: the PHP C runtime, Guzzle’s transport layer, and Google’s SDK adapters. What used to be impossible in standard PHP came together in 2025–2026 across three key milestones:</p><p><strong>Layer:</strong> <strong>PHP 8.5</strong><br><strong>Release:</strong> Nov 2025<br><strong>What Changed:</strong> Added curl_share_init_persistent(), allowing underlying cURL connection pools to survive across requests inside the worker process instead of getting destroyed at request shutdown.</p><p><strong>Layer:</strong> <strong>Guzzle 8.0 &amp; 8.2</strong><br><strong>Release:</strong> Jul–Sep 2026<br><strong>What Changed:</strong> Added TransportSharing::PERSISTENT_PREFER (closing issue <a href="https://github.com/guzzle/guzzle/issues/3307">#3307</a>), native HTTP/2 multiplexing, and automatic connection reuse.</p><p><strong>Layer:</strong> <strong>Google Cloud SDK</strong><br><strong>Release:</strong> Mid–Late 2026<br><strong>What Changed:</strong> Updated google/auth and google/cloud-core with official Guzzle 8 support and shared handler pooling.</p><h3>Why curl_reset() Doesn&#39;t Close Your Sockets</h3><p>A common question from engineers looking at Guzzle’s internals is:</p><blockquote>“Guzzle resets its cURL easy handles and tears them down at request shutdown. Doesn’t that close the connection?”</blockquote><p>Under standard PHP, yes. But with PHP 8.5’s persistent sharing enabled, <strong>libcurl moves socket ownership out of the individual request handle and into the worker process’s shared connection pool</strong>.</p><p>When a request finishes, only the temporary easy handle is reset. The underlying TCP socket and TLS 1.3 session remain safely parked in the worker process. When the next request arrives, Guzzle simply re-attaches a new handle to the warm socket — giving you immediate socket reuse with 0ms connection overhead.</p><h3>Two Critical Gotchas in Guzzle 8.2</h3><p>Before wiring this into production, there are two non-obvious traps in Guzzle 8.2 that can quietly disable your socket pooling:</p><h4>Gotcha #1: Multiplexing Is Silently Disabled Unless You Set &#39;version&#39; =&gt; 2.0</h4><p>Guzzle 8 enables HTTP/2 multiplexing (CURLOPT_PIPEWAIT) by default. This tells cURL to wait and reuse an existing warm HTTP/2 connection instead of opening parallel TCP sockets.</p><p>However, PSR-7 requests default to HTTP/1.1. If you don’t explicitly configure HTTP/2 on the Guzzle client, Guzzle silently falls back to HTTP/1.1 — completely disabling stream multiplexing.</p><ul><li><strong>The fix:</strong> Always set &#39;version&#39; =&gt; 2.0 in your top-level client options:</li></ul><pre>$client = new GuzzleHttp\Client([<br>      &#39;version&#39;           =&gt; 2.0, // Enables HTTP/2 &amp; pipewait multiplexing<br>      &#39;transport_sharing&#39; =&gt; TransportSharing::PERSISTENT_PREFER,<br>  ]);</pre><h4>Gotcha #2: Don’t Set Connection Limits on Standard Linux Images (libcurl &lt; 8.22.0)</h4><p>Guzzle 8.2 allows setting max_host_connections to cap pooled sockets. However, due to a bug in older libcurl versions, Guzzle will <strong>silently disable persistent sharing</strong> if connection caps are set on libcurl &lt; 8.22.0.</p><p>Most Debian and Ubuntu base images currently ship with libcurl 8.12.x. Setting max_host_connections on these images will quietly turn off cross-request socket pooling without throwing an error.</p><ul><li><strong>The fix:</strong> Leave max_host_connections unset unless your base container explicitly runs libcurl &gt;= 8.22.0.</li></ul><h3>The Implementation: Wiring Persistent Transport Sharing Across Google Cloud SDKs</h3><p>To ensure every outbound call in a Cloud Run worker — whether made by BigQueryClient, DatastoreClient, ApplicationDefaultCredentials, or custom async Guzzle calls—shares the exact same persistent socket pool and HTTP/2 multiplexed streams, we define a centralized factory in our bootstrap configuration:</p><pre>&lt;?php<br> <br>use Google\Auth\Cache\SysVCacheItemPool;<br>use Google\Auth\CredentialsLoader;<br>use Google\Auth\HttpHandler\HttpHandlerFactory;<br>use Google\Cloud\BigQuery\BigQueryClient;<br>use Google\Cloud\Datastore\DatastoreClient;<br>use GuzzleHttp\Client;<br>use GuzzleHttp\TransportSharing;<br> <br>/**<br> * Returns a shared Guzzle 8.2 Client backed by PHP 8.5&#39;s curl_share_init_persistent()<br> * so TCP sockets, TLS 1.3 sessions, and DNS lookups to *.googleapis.com stay warm<br> * across sync/async handlers, across SDK clients, and across Cloud Run requests.<br> */<br>function getSharedGuzzleClient(array $overrides = []): Client<br>{<br>    static $defaultClient = null;<br>    if ($defaultClient !== null &amp;&amp; empty($overrides)) {<br>        return $defaultClient;<br>    }<br> <br>    $config = [<br>        &#39;version&#39;           =&gt; 2.0, // Unlocks HTTP/2 + Guzzle 8.2 Multiplexing::WAIT (CURLOPT_PIPEWAIT)<br>        &#39;timeout&#39;           =&gt; 30,<br>        &#39;connect_timeout&#39;   =&gt; 5,<br>        &#39;decode_content&#39;    =&gt; &#39;gzip&#39;,<br>        &#39;headers&#39;           =&gt; [<br>            &#39;Accept-Encoding&#39; =&gt; &#39;gzip&#39;,<br>            &#39;User-Agent&#39;      =&gt; &#39;cloud-run-analytics-api/2.0 (gzip)&#39;,<br>        ],<br>        // The magic switch in Guzzle 8.0+ / 8.2.0:<br>        // Attaches CurlSharePersistentHandle(DNS | CONNECT | SSL_SESSION) to both<br>        // CurlHandler (sync) and CurlMultiHandler (async)<br>        &#39;transport_sharing&#39; =&gt; TransportSharing::PERSISTENT_PREFER,<br>        &#39;curl&#39;              =&gt; [<br>            CURLOPT_TCP_KEEPALIVE =&gt; 1,<br>            CURLOPT_TCP_KEEPIDLE  =&gt; 60,<br>            CURLOPT_TCP_KEEPINTVL =&gt; 30,<br>        ],<br>    ];<br> <br>    if (!empty($overrides)) {<br>        return new Client(array_replace_recursive($config, $overrides));<br>    }<br> <br>    return $defaultClient = new Client($config);<br>}<br> <br>/**<br> * Wraps the shared Guzzle 8.2 client in Google Auth&#39;s HTTP handler.<br> * Requires google/auth &gt;= 1.53.0 (latest v1.55.2).<br> */<br>function getSharedGoogleHttpHandler()<br>{<br>    static $handler = null;<br>    if ($handler === null) {<br>        $handler = HttpHandlerFactory::build(getSharedGuzzleClient());<br>    }<br>    return $handler;<br>}<br> <br>// Share both the data transport AND auth token transport across all Google Cloud clients:<br>$sharedHttpHandler = getSharedGoogleHttpHandler();<br>$sysvAuthCache     = new SysVCacheItemPool();<br> <br>$bigQuery = new BigQueryClient([<br>    &#39;projectId&#39;       =&gt; $_ENV[&#39;PROJECT_ID&#39;],<br>    &#39;httpHandler&#39;     =&gt; $sharedHttpHandler,<br>    &#39;authHttpHandler&#39; =&gt; $sharedHttpHandler,<br>    &#39;authCache&#39;       =&gt; $sysvAuthCache,<br>]);<br> <br>$datastore = new DatastoreClient([<br>    &#39;projectId&#39;       =&gt; $_ENV[&#39;PROJECT_ID&#39;],<br>    &#39;httpHandler&#39;     =&gt; $sharedHttpHandler,<br>    &#39;authHttpHandler&#39; =&gt; $sharedHttpHandler,<br>    &#39;authCache&#39;       =&gt; $sysvAuthCache,<br>]);</pre><h3>Exposing Live Socket Reuse Telemetry via on_stats Response Headers</h3><p>To verify socket reuse directly inside Postman or curl without digging through container logs, you can attach an on_stats callback to $config when a debug flag is present. This inspects libcurl&#39;s microsecond counters (namelookup_time_us, connect_time_us, appconnect_time_us) and emits the exact X-Transport-Trace-* headers shown at the start of this article:</p><pre>$config[&#39;on_stats&#39;] = function (\GuzzleHttp\TransferStats $stats) {<br>    if (headers_sent()) {<br>        return;<br>    }<br>    static $callIndex = 0;<br>    $callIndex++;<br>    $s       = $stats-&gt;getHandlerStats();<br>    $host    = $stats-&gt;getRequest()-&gt;getUri()-&gt;getHost();<br>    $port    = $s[&#39;local_port&#39;] ?? 0;<br>    $dnsMs   = round(($s[&#39;namelookup_time_us&#39;] ?? 0) / 1000, 2);<br>    $tcpMs   = round((($s[&#39;connect_time_us&#39;] ?? 0) - ($s[&#39;namelookup_time_us&#39;] ?? 0)) / 1000, 2);<br>    $tlsMs   = ($s[&#39;appconnect_time_us&#39;] ?? 0) &gt; 0<br>        ? round((($s[&#39;appconnect_time_us&#39;] ?? 0) - ($s[&#39;connect_time_us&#39;] ?? 0)) / 1000, 2)<br>        : 0.0;<br>    $ttfbMs  = round(($s[&#39;starttransfer_time_us&#39;] ?? 0) / 1000, 2);<br>    $totalMs = round(($s[&#39;total_time_us&#39;] ?? 0) / 1000, 2);<br>    $reused  = ($s[&#39;connect_time_us&#39;] ?? 0) === 0 ? &#39;YES (0ms Handshake)&#39; : &#39;NO (Cold Handshake)&#39;;<br> <br>    header(&#39;X-Guzzle-Transport: Guzzle &#39; . \GuzzleHttp\ClientInterface::MAJOR_VERSION . &#39; | mode=persistent_prefer | pid=&#39; . getmypid());<br>    header(sprintf(<br>        &#39;X-Transport-Trace-%d: host=%s local_port=%d dns=%.2fms tcp=%.2fms tls=%.2fms ttfb=%.2fms total=%.2fms socket_reused=%s&#39;,<br>        $callIndex,<br>        $host,<br>        $port,<br>        max(0, $dnsMs),<br>        max(0, $tcpMs),<br>        max(0, $tlsMs),<br>        $ttfbMs,<br>        $totalMs,<br>        $reused<br>    ));<br>};</pre><h3>Benchmarking: Sockets Are Pooled Per Worker Process (pid)</h3><p>When testing against Cloud Run in Postman or cURL, remember that connection pools live in worker process memory.</p><p>In containers running PHP-FPM or Apache prefork, incoming requests round-robin across a pool of child worker processes:</p><ul><li><strong>Initial requests hit cold workers:</strong> If your container has 8 workers, your first few clicks in Postman will likely route to different workers, showing cold handshakes on each one.</li><li><strong>Warm reuse:</strong> Once a worker has handled its first request, subsequent requests landing on that same process will immediately show 0.00ms handshakes and socket reuse.</li><li><strong>Avoid idle socket drops:</strong> Google’s frontend closes idle HTTP/2 connections after 60–120 seconds. Align your PHP-FPM worker count with your Cloud Run container concurrency so active requests keep warm workers in regular rotation.</li></ul><p>Cold Worker — Opens Sockets &amp; Seeds CurlSharePersistentHandle</p><pre>X-Transport-Trace-1: host=oauth2.googleapis.com local_port=51412 dns=34.33ms tcp=27.03ms tls=92.22ms ttfb=202.62ms total=203.27ms socket_reused=NO (Cold Handshake)<br>X-Transport-Trace-2: host=bigquery.googleapis.com local_port=51415 dns=32.43ms tcp=57.01ms tls=175.31ms ttfb=1653.11ms total=1667.24ms socket_reused=NO (Cold Handshake)<br>X-Transport-Trace-3: host=datastore.googleapis.com local_port=51429 dns=72.66ms tcp=50.61ms tls=153.86ms ttfb=1409.00ms total=1409.23ms socket_reused=NO (Cold Handshake)</pre><p>Warm Worker — Reuses All 3 Sockets with 0.00ms Handshakes</p><pre>X-Transport-Trace-1: host=oauth2.googleapis.com local_port=51412 dns=0.07ms tcp=0.00ms tls=0.00ms ttfb=55.57ms total=56.05ms socket_reused=YES (0ms Handshake)<br>X-Transport-Trace-2: host=bigquery.googleapis.com local_port=51415 dns=0.02ms tcp=0.00ms tls=0.00ms ttfb=632.90ms total=652.75ms socket_reused=YES (0ms Handshake)<br>X-Transport-Trace-3: host=datastore.googleapis.com local_port=51429 dns=0.03ms tcp=0.00ms tls=0.00ms ttfb=417.27ms total=417.50ms socket_reused=YES (0ms Handshake)</pre><h3>What Else Changed in the Google Cloud PHP SDKs</h3><p>Upgrading to the latest Google Cloud SDK packages also brings three important latency and stability improvements:</p><h4>1. BigQuery Runs Stateless Short Queries by Default</h4><p>Following up on my previous article on <a href="https://medium.com/google-cloud/how-we-halved-latency-in-php-with-bigquery-short-query-mode-c29eee204b0a">Halving Latency with BigQuery Short Query Mode</a>, standard queries run through BigQueryClient::runQuery() now execute in stateless short-query mode out of the box—skipping job creation overhead entirely.</p><blockquote><strong><em>Pagination Watchout:</em></strong><em> Because fast queries complete without creating a background Job, </em><em>$queryResults-&gt;info()[&#39;jobReference&#39;] is now </em><em>null. If your pagination relies on a materialized destination table, chain </em><em>-&gt;createDisposition(&#39;CREATE_IF_NEEDED&#39;) onto your query to force stateful execution.</em></blockquote><h4>2. Full-TTL OAuth2 Token Caching</h4><p>google/auth now respects the OAuth2 token&#39;s full 1-hour expiration window when storing tokens in shared process caches (SysVCacheItemPool), eliminating premature token refreshes across PHP-FPM workers.</p><h4>3. No More Double-Fetched Pages</h4><p>google/cloud-core resolved an issue where iterating paginated REST results could accidentally re-fetch the first page over the network.</p><h3>Conclusion: A Must-Have Upgrade for Cloud Run PHP Services</h3><p>For years, serverless PHP accepted a 50ms–150ms connection tax on every single API call to BigQuery, Datastore, or Vertex AI.</p><p>With PHP 8.5’s persistent cURL handles, Guzzle 8.2’s transport sharing, and modern Google Cloud SDKs, standard Cloud Run containers finally achieve the same zero-overhead HTTP/2 multiplexing as long-lived Go or Java services — with a simple 15-line configuration change.</p><h3>Wrap Up</h3><p>In the meantime, check it out other contents from me:</p><ul><li><a href="https://medium.com/google-cloud/are-native-php-binaries-on-serverless-for-real-deployed-it-to-cloud-run-here-are-the-results-d994e85bd359">Are Native PHP Binaries on Serverless for Real? Deployed It to Cloud Run — Here Are the Results</a></li><li><a href="https://medium.com/google-cloud/how-we-halved-latency-in-php-with-bigquery-short-query-mode-c29eee204b0a">How We Halved Latency in PHP with BigQuery Short Query Mode</a></li><li><a href="https://medium.com/p/195eb98c3d02">I Flipped a Switch in BigQuery and My Queries Got 30% Faster for Free</a></li><li><a href="https://medium.com/p/9e2634c84a82">Reduce your BigQuery bills with BI Engine capacity orchestration</a></li><li><a href="https://medium.com/p/def058a7c8dc">Hacking ADK’s Importer: How We Slashed 24-Second Cold Start in Half</a></li><li><a href="https://medium.com/p/cb03976f47c1">Introducing “GCP Project Switcher” Extension for VS Code — My Weekend Project</a></li><li><a href="https://medium.com/p/64da29fd58d1">Integrating DockerSlim container minify step on Cloud Build</a></li></ul><p>Feel free to reach out to me on <a href="https://www.linkedin.com/in/pentium10/">Linkedin</a> or read my previous posts on <a href="https://martonkodok.medium.com/">medium/@martonkodok</a></p><p><em>We at </em><a href="https://reea.net/"><em>REEA</em></a><em> are open to providing specialized knowledge, software development to your Google Cloud needs. Get in contact with Marton to kickstart discussions.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ff749fd02a9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/php-on-cloud-run-cut-your-latency-with-php-8-5-and-persistent-sockets-7ff749fd02a9">PHP on Cloud Run: Cut Your Latency with PHP 8.5 and Persistent Sockets</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Are Native PHP Binaries on Serverless for Real? Deployed It to Cloud Run — Here Are the Results]]></title>
            <link>https://medium.com/google-cloud/are-native-php-binaries-on-serverless-for-real-deployed-it-to-cloud-run-here-are-the-results-d994e85bd359?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/d994e85bd359</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[serverless]]></category>
            <category><![CDATA[php]]></category>
            <category><![CDATA[typephp]]></category>
            <category><![CDATA[google-cloud-run]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 08:04:15 GMT</pubDate>
            <atom:updated>2026-10-18T10:52:33.161Z</atom:updated>
            <content:encoded><![CDATA[<h3>Are Native PHP Binaries on Serverless for Real? Deployed It to Cloud Run — Here Are the Results</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3Ue4vmsld0vXCmMcNuIKcg.png" /></figure><p><em>When the Swoole team announced </em><a href="https://github.com/swoole/typephp"><em>TypePHP</em></a><em>, I had to test it on Google Cloud Run to see if compiling down to a single binary is truly that simple to containerize, and whether the performance gains are real without rewriting code in Rust or Go.</em></p><h3>What is TypePHP?</h3><p>Swoole’s open-source <a href="https://github.com/swoole/typephp"><strong>TypePHP</strong></a> an Ahead-Of-Time (AOT) compiler that translates PHP source code directly into C++ and compiles it into native standalone machine code.</p><p>Unlike traditional PHP runtimes or opcode caches, TypePHP doesn’t interpret opcodes at runtime — it compiles PHP functions directly into optimized native CPU binaries. For more details on the compiler internals, check out the <a href="https://github.com/swoole/typephp">TypePHP GitHub repository</a> and their <a href="https://swoole.com/aot/en">official documentation</a>.</p><p>As a Cloud Architect working extensively with serverless platforms, my immediate thought was:</p><blockquote><strong>“Can we pair TypePHP with Google Cloud Run? Is it actually that simple to package inside a Dockerfile? And are the near-native performance gains and microsecond response times real when deployed to a live serverless environment?”</strong></blockquote><p>I decided to put it to the test by migrating an existing real-world PHP workload onto the TypePHP / standalone runtime on <strong>Google Cloud Run</strong>.</p><p>Here is what happened, how your existing application gets transformed during the build, and the benchmark numbers we recorded.</p><h3>The Serverless Context: PHP is Fast, But We Wanted More</h3><p>To be fair, PHP has always been surprisingly agile in serverless environments.</p><p>With lightweight base runtimes, PHP cold starts typically land comfortably in the <strong>300ms to 700ms range</strong> — a stark contrast to the sluggish cold starts often observed in the Python or JVM ecosystems, where loading heavy modules (like Pandas, NumPy, or Spring Boot) can drag cold starts out anywhere from <strong>3 to 27+ seconds</strong>.</p><p>However, once you operate on serverless containers, you are naturally driven to chase every last drop of optimization: <a href="https://medium.com/google-cloud/integrating-dockerslim-container-minify-step-on-cloud-build-64da29fd58d1">minify with DockerSlim</a>, eliminating container daemon bloat (Nginx + PHP-FPM + supervisord), and squeezing maximum concurrency out of every single vCPU.</p><p>TypePHP and its static standalone ecosystem (swoole-cli / tpc) allow us to push past the &quot;fast enough&quot; baseline and turn PHP into an ultra-lean, microsecond in-memory powerhouse.</p><h3>What Enables the Trick: The “Single Static Binary Image”</h3><p>Think about how Go or Rust applications are deployed: you don’t ship a directory full of loose source files and dynamic interpreters to production. You compile everything Ahead-Of-Time into a single, standalone binary artifact.</p><p><strong>What if you could achieve the exact same thing with PHP at the build phase?</strong></p><p>Instead of hosting a directory with hundreds of PHP files, templates, and dependencies — forcing the server to parse scripts and hit disk — the application is compiled Ahead-Of-Time into a single standalone binary.</p><h3>The Build-to-Cloud Pipeline:</h3><ol><li><strong>Build Stage<br></strong>TypePHP compiles your PHP application Ahead-Of-Time (AOT) and packages all application logic into an immutable, self-contained static <strong>binary image</strong>.</li><li><strong>Runtime on Cloud Run:</strong> The container boots in under 700ms. A single binary listens directly on $PORT delivering sub-millisecond execution (&lt; 1.5ms TTFB) with <strong>zero disk I/O</strong> during live traffic.</li></ol><p>If you have a use case like a static website, a background job script, a data processing worker, or any task that is packagable into a single binary file, you are now operating with near-native edge performance — <strong>without having to rewrite a single line of your existing PHP codebase into Rust, Go, or C++</strong>.</p><p>You simply ship a binary, eliminate Nginx/PHP-FPM glue, and let Cloud Run serve requests at memory speed.</p><h3>Getting Started: The Composer Route</h3><p>Before jumping into the container build, it is worth noting how easy TypePHP is to integrate into everyday PHP workflows. You can pull the compiler directly into your project via Composer:</p><pre>composer require swoole/typephp</pre><p>Once installed, compile and run it:</p><pre>vendor/bin/tpc.php hello.php -O2 -o hello<br>./hello</pre><p>Under the hood, TypePHP leverages <strong>PHPx</strong> — Swoole’s C++ abstraction layer that seamlessly bridges compiled native C++ code with the Zend VM runtime, allowing your compiled functions to interoperate with dynamic PHP functions and standard extensions. For step-by-step instructions on making the <strong>PHPx runtime and embed libraries available</strong> on your system, check out the <a href="https://swoole.com/aot/en/docs/composer">official TypePHP Composer documentation</a>.</p><h3>What Actually Happens During the Build Step?</h3><p>You might wonder: <em>How does my existing PHP code get transformed into an in-memory execution engine during </em><em>docker build?</em></p><p>Here is the transformation that takes place during a build step:</p><p>• <strong>Automatic Source &amp; Asset Discovery:</strong> The build runner scans your project directory, discovering all templates, dynamic route scripts, metadata files, and static assets without requiring manual configuration.<br>• <strong>Pre-compilation into In-Memory Registry:</strong> The build step renders and packages your canonical routes and static files into an optimized in-memory registry. In production, this registry resides directly in RAM as an $O(1)$ lookup table.<br>• <strong>Zero-Disk Runtime Shift:</strong> When the container runs on Cloud Run, the router handles incoming requests directly from RAM through Swoole coroutines. The traditional filesystem traversal (reading .php files off disk on every hit) is completely bypassed.</p><h3>The Dockerfile: Packaging &amp; Serving the Native Binary</h3><p>Here is the complete multi-stage Dockerfile showing how to compile your application with TypePHP and serve the standalone app-server binary in a minimal runner container:</p><pre># =============================================================================<br># Stage 1: Build &amp; Compile PHP to Native Binary using TypePHP (tpc)<br># =============================================================================<br>FROM debian:bookworm-slim AS builder<br>RUN apt-get update &amp;&amp; apt-get install -y --no-install-recommends \<br>    gcc g++ cmake make git curl ca-certificates libphp8.4-embed libgmp-dev<br>WORKDIR /app<br>COPY . /app<br># Compile PHP application down to native AOT binary with TypePHP<br>RUN composer require swoole/typephp &amp;&amp; \<br>    vendor/bin/tpc.php server.php -O3 -o /app/build/app-server<br># =============================================================================<br># Stage 2<br># =============================================================================<br>FROM debian:bookworm-slim AS runner<br>RUN apt-get update &amp;&amp; apt-get install -y --no-install-recommends \<br>    libphp8.4-embed ca-certificates &amp;&amp; \<br>    rm -rf /var/lib/apt/lists/*<br># Copy phpx runtime shared library from builder<br>COPY --from=builder /usr/lib/libphpx.so /usr/lib/libphpx.so<br>RUN ldconfig<br>WORKDIR /app<br>COPY --from=builder /app/build/app-server /app/app-server<br># Direct native binary execution (Instant sub-millisecond boot)<br>EXPOSE 8080<br>ENTRYPOINT [&quot;/app/app-serverdoc&quot;, &quot;--port=8080&quot;, &quot;--host=0.0.0.0&quot;]</pre><h3>Deploying to Cloud Run in Seconds</h3><p>Deploying the final container image to Google Cloud Run is a single command:</p><pre>gcloud run deploy my-typephp-service \<br>  --image=gcr.io/my-project/my-typephp-app \<br>  --region=us-central-1 \<br>  --platform=managed \<br>  --allow-unauthenticated \<br>  --concurrency=250 \<br>  --cpu=1 \<br>  --memory=128Mi</pre><h3>The Results: Are the Performance Gains Real?</h3><p>We deployed the container to Google Cloud Run and ran comprehensive load tests comparing the traditional dynamic PHP stack against the TypePHP standalone container under identical resources (1 vCPU, 256MB RAM):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/778/1*DtP-wDaIrIxruEhD8lHmqg.png" /><figcaption>Request Time Log on Cloud Run</figcaption></figure><p>• <strong>Container Image Size:</strong> Dropped from ~350 MB down to <strong>&lt; 35 MB</strong> (90% reduction in image weight).<br>• <strong>Container Boot Time:</strong> Reduced from ~2.5s down to <strong>&lt; 150–500 ms</strong> (15–10x faster cold-start wakeups).<br>• <strong>Time to First Byte (TTFB):</strong> Plummeted from 45ms–180ms down to <strong>&lt; 1.5 ms</strong> (30x to 100x faster responses).<br>• <strong>Throughput per Container:</strong> Surged from ~180 RPS to <strong>2,800+ RPS</strong> on a single vCPU (15.5x increase in concurrency).<br>• <strong>Live Request Disk I/O:</strong> Reduced from 25–40 IOPS to <strong>0 IOPS</strong> (100% in-memory execution).<br>• <strong>Network Bandwidth:</strong> Reduced by over <strong>80%</strong> via automatic on-the-fly compression and 304 ETag caching.</p><p>The verdict is unequivocal: <strong>the performance gains are real.</strong> Eliminating disk I/O and process-per-request overhead on Cloud Run drops server response times to microseconds and makes cold starts practically unnoticeable.</p><h3>Recommended Use Cases: Where This Architecture Shines</h3><ol><li><strong>Static Websites &amp; Developer Documentation:</strong> Sites and documentation engines that can be compiled down to a single binary. Every page and asset is served directly from RAM without runtime disk reads.</li><li><strong>Standalone CLI Background Jobs &amp; Batch Processors:</strong> Queue workers, ETL scripts, or offline batch calculators that can be compiled into a single binary and triggered via Cloud Run Jobs with zero cold-start delay.</li><li><strong>High-Throughput Read-Heavy APIs &amp; Gateways:</strong> Microservices serving configuration dictionaries, feature flags, localization strings, or public catalogs where predictable sub-millisecond responses are mandatory.</li><li><strong>Cloud Run &amp; Serverless Burst-Traffic APIs:</strong> Workloads requiring instant scaling without cold-start CPU penalties or worker exhaustion.</li><li><strong>Edge Containers &amp; Kiosks:</strong> Low-memory devices, IoT dashboards, and offline-capable edge runtimes running smoothly inside a 35MB footprint.</li></ol><h3>Conclusion</h3><p>TypePHP it promises near-native C++ performance for PHP applications.</p><p>Deploying it to <strong>Google Cloud Run</strong> proved that not only is the setup remarkably straightforward via modern Dockerfiles, but compiling down to a single static binary gives you the execution speed and lightweight footprint of Go or Rust — without leaving the PHP ecosystem you already know and love.</p><h3>Wrap Up</h3><p>In the meantime, check it out other contents from me:</p><ul><li><a href="https://martonkodok.medium.com/php-on-cloud-run-cut-your-latency-with-php-8-5-and-persistent-sockets-7ff749fd02a9">PHP Devs on Cloud Run: Cut Your Latency with PHP 8.5 and Guzzle 8.2 Persistent Sockets</a></li><li><a href="https://medium.com/google-cloud/how-we-halved-latency-in-php-with-bigquery-short-query-mode-c29eee204b0a">How We Halved Latency in PHP with BigQuery Short Query Mode</a></li><li><a href="https://medium.com/p/195eb98c3d02">I Flipped a Switch in BigQuery and My Queries Got 30% Faster for Free</a></li><li><a href="https://medium.com/p/9e2634c84a82">Reduce your BigQuery bills with BI Engine capacity orchestration</a></li><li><a href="https://medium.com/p/def058a7c8dc">Hacking ADK’s Importer: How We Slashed 24-Second Cold Start in Half</a></li><li><a href="https://medium.com/p/cb03976f47c1">Introducing “GCP Project Switcher” Extension for VS Code — My Weekend Project</a></li><li><a href="https://medium.com/p/64da29fd58d1">Integrating DockerSlim container minify step on Cloud Build</a></li></ul><p>Feel free to reach out to me on <a href="https://www.linkedin.com/in/pentium10/">Linkedin</a> or read my previous posts on <a href="https://martonkodok.medium.com/">medium/@martonkodok</a></p><p><em>We at </em><a href="https://reea.net/"><em>REEA</em></a><em> are open to providing specialized knowledge, software development to your Google Cloud needs. Get in contact with Marton to kickstart discussions.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d994e85bd359" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/are-native-php-binaries-on-serverless-for-real-deployed-it-to-cloud-run-here-are-the-results-d994e85bd359">Are Native PHP Binaries on Serverless for Real? Deployed It to Cloud Run — Here Are the Results</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cross-Agent Synergy: How I keep Gemini CLI in sync with other AI tools?]]></title>
            <link>https://medium.com/google-cloud/cross-agent-synergy-how-i-keep-gemini-cli-in-sync-with-other-ai-tools-ae5ae153a0c6?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/ae5ae153a0c6</guid>
            <category><![CDATA[github-copilot]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[gemini-cli]]></category>
            <category><![CDATA[skills]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Wed, 11 Mar 2026 15:44:15 GMT</pubDate>
            <atom:updated>2026-03-12T12:07:17.965Z</atom:updated>
            <content:encoded><![CDATA[<p>In today’s fast-moving development world, different teams often rely on distinct solutions like <strong>Gemini CLI</strong>, <strong>VS Code Copilot</strong> and <strong>Cursor</strong>, meaning “the AI” is rarely just one tool.</p><blockquote><em>Personally, I prefer </em><a href="https://geminicli.com/"><strong><em>Gemini CLI</em></strong></a><em> over other tools — </em><strong><em>I identify more with its prompting style and the high quality of code it produces</em></strong><em> — though my human colleagues often switch between these various platforms for surgical, programmatic power.</em></blockquote><p>But each tool has its “own way” of defining and use expertise. In these, we found .agents/skills folder the universal way available across all tool, containing specialized instructions (SKILL.md). This gives me a unified, markdown-driven logic that works everywhere, keeping me synced while staying ahead of the curve — I am the one leading top-notch solutions to the team.</p><blockquote><em>While Gemini CLI is powerful enough to discover and use these </em><em>skill files automatically during its reasoning process. A feature I like in Gemini CLI, we could rely on</em> <em>command files in</em> <em>.gemini/command folder available as /slash prompts.</em></blockquote><p>I prefer having them explicitly available as <strong>Slash Commands</strong>. This allows for a more “surgical” approach, where I can invoke a specific expert skill instantly without relying on the model’s general discovery.</p><p>Here’s how I automated the <strong>linking </strong>of these skills into my native Gemini CLI commands.</p><h3>The Transformation</h3><p><strong>Before:</strong></p><pre>.<br>├── .agents/<br>│   └── skills/<br>│       └── agent-generate-agent-md/<br>│           └── SKILL.md<br>└── .gemini/<br>    └── settings.json</pre><p><strong>After:</strong></p><pre>.<br>├── .agents/<br>│   └── skills/<br>│       └── agent-generate-agent-md/<br>│           └── SKILL.md<br>└── .gemini/<br>    ├── commands/<br>    │   └── agentGenerateAgentMd.toml<br>    └── settings.json</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PDnhUtgPz2qJ5hxwOP9yxA.png" /></figure><h3>The Shortcut: /agentSyncSkills</h3><p>One feature I really love in the Gemini CLI is the ability to run commands with slash /myCommand. It’s the fastest way to call a specialized skill with custom parameters. It turns a static prompt into a dynamic, reusable tool.</p><p>However, when my human collegues adds a new skill to the .agents folder, I’m not automatically aware of it as a Gemini command.</p><h3>So, I built a command for myself: /agentSyncSkills.</h3><p>To automate this bridge, I use a custom command definition. This command tells me how to scan the project for external skills and transform them into native Gemini CLI commands.</p><pre>description=&quot;sync skills from .agents folder&quot;<br>prompt = &#39;&#39;&#39;<br>The goal of this prompt is to create custom commands for the Gemini CLI <br>from a universally defined skills folder.<br><br> You need to search the existing folders from `.agents/skills/`.<br>Each folder name describes an agent. Inside each folder is a `SKILL.md` file.<br><br>Your job is to create a `.gemini/commands/agent[Name].toml` file for each new skill you find.<br>Rules for the generated files:<br><br>1. Naming convention: each file starts with the &quot;agent&quot; prefix, followed by PascalCase (e.g., `agentSyncSkills.toml`).<br>2. The commands must reference the skills file rather than reusing the prompt contents.<br>3. You must use the EXACT path to the `SKILL.md` file in the `@` include syntax.<br><br>Below is the template for the files you need to generate. <br><br>CRITICAL INSTRUCTION: In the template below, I have used `[AT]` to represent the &quot;at&quot; symbol. <br>When you generate the final TOML files in the bash script, you MUST replace `[AT]` with the actual `@` character so the file injection works.<br><br>---<br>description=&quot;you put here a brief description of the agent&quot;<br>prompt = &quot;&quot;&quot;<br>Use the following files instructions to resolve the user&#39;s ask:<br>[AT]{.agents/skills/[InsertFolderNameHere]/SKILL.md}<br>&quot;&quot;&quot;<br>---<br><br>No need to create a global search for SKILL.md files to ensure no other skills were missed.<br>Output the list of commands you just created. Do not create files for skills that already have a .toml file.<br><br>And inform me to run /commands reload<br>&#39;&#39;&#39;<br></pre><h3>Automating the Sync</h3><p>My /agentSyncSkills workflow follows a simple but powerful logic:</p><ol><li><strong>Discovery:</strong> it lists the contents of .agents/skills/ and compare it with .gemini/commands/.</li><li><strong>Extraction:</strong> For any missing command, it reads the corresponding SKILL.md. I extract its description and instructions.</li><li><strong>Generation:</strong> creates a new .toml file to .gemini/commands/, mapping the skill’s logic into our custom command format.</li></ol><p>Now, whenever a new skill is added to the project, I can “digest” it and make it available as a /agentMyCommand shortcut in seconds.</p><blockquote><strong><em>NOTE</em></strong><em>: As you see I use </em><strong><em>references </em></strong><em>in my Gemini CLI command definitions @</em><em>{.agents/skills/[InsertFolderNameHere]/SKILL.md}</em> <strong><em>which gives me the leverage as a symlink</em></strong><em>, so whenever the skill is modified my command does not need to be updated.</em></blockquote><h3>The Missing Link: Programmatic Reloading</h3><p>There’s one final hurdle in this automation. After I generate these new command files, I need to manually run /commands reload.</p><p><strong>Can an agent instruct the CLI to reload itself?</strong></p><p>Currently, the answer is <strong>no</strong>. Slash commands like /commands reload are user-level meta-controls.</p><p>My settings.json</p><pre>{<br>  &quot;ui&quot;: {<br>    &quot;hideBanner&quot;: true<br>  },<br>  &quot;context&quot;: {<br>    &quot;fileName&quot;: [&quot;AGENTS.md&quot;]<br>  }<br>}</pre><h3>Wrap Up</h3><p>In the meantime, check it out other contents from me:</p><ul><li><a href="https://medium.com/google-cloud/how-we-halved-latency-in-php-with-bigquery-short-query-mode-c29eee204b0a">How We Halved Latency in PHP with BigQuery Short Query Mode</a></li><li><a href="https://medium.com/p/195eb98c3d02">I Flipped a Switch in BigQuery and My Queries Got 30% Faster for Free</a></li><li><a href="https://medium.com/p/9e2634c84a82">Reduce your BigQuery bills with BI Engine capacity orchestration</a></li><li><a href="https://medium.com/p/def058a7c8dc">Hacking ADK’s Importer: How We Slashed 24-Second Cold Start in Half</a></li><li><a href="https://medium.com/p/cb03976f47c1">Introducing “GCP Project Switcher” Extension for VS Code — My Weekend Project</a></li><li><a href="https://medium.com/p/64da29fd58d1">Integrating DockerSlim container minify step on Cloud Build</a></li></ul><p>Feel free to reach out to me on <a href="https://www.linkedin.com/in/pentium10/">Linkedin</a> or read my previous posts on <a href="https://martonkodok.medium.com/">medium/@martonkodok</a></p><p><em>We at </em><a href="https://reea.net/"><em>REEA</em></a><em> are open to providing specialized knowledge, software development to your Google Cloud needs. Get in contact with Marton to kickstart discussions.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ae5ae153a0c6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/cross-agent-synergy-how-i-keep-gemini-cli-in-sync-with-other-ai-tools-ae5ae153a0c6">Cross-Agent Synergy: How I keep Gemini CLI in sync with other AI tools?</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How We Halved Latency in PHP with BigQuery Short Query Mode]]></title>
            <link>https://medium.com/google-cloud/how-we-halved-latency-in-php-with-bigquery-short-query-mode-c29eee204b0a?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/c29eee204b0a</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[php]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[bigquery]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Mon, 09 Feb 2026 21:08:35 GMT</pubDate>
            <atom:updated>2026-02-10T10:30:15.489Z</atom:updated>
            <content:encoded><![CDATA[<p>In the world of web development, especially with “snap-on” PHP applications, latency is the enemy. We extensively use Google BigQuery for our data analytics and search features. While BigQuery is a powerhouse for processing massive datasets, we noticed a significant bottleneck when running small, frequent queries — like returning small amounts of rows from a table. <a href="https://martonkodok.medium.com/c29eee204b0a?source=friends_link&amp;sk=df93ed71c010a87096523a2521c3d4db"><em>Friendly reading link</em></a>.</p><blockquote>We managed to cut our query latency by <strong>50% (and often more)</strong> by switching to BigQuery’s <strong>Short Query Mode</strong> (<em>jobCreationMode=JOB_CREATION_OPTIONAL</em>).</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8uXljzm4F1bCQE9ERs95Pg.png" /></figure><h3>The Problem: The “Job” Overhead</h3><p>The official Google Cloud PHP library for BigQuery is robust, but it defaults to a specific workflow for running queries: it creates a <strong>Job</strong>.</p><p>Every time you run $bigQuery-&gt;query(&#39;SELECT ...&#39;), the client:</p><ol><li>Sends a request to create a Query Job.</li><li>Polls the server to check if the Job is complete.</li><li>Retrieves the results once finished.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*UouUFwF4zm5QIFCCr88wTg.png" /></figure><p>For analyzing terabytes of data, this is perfect. But for a simple query like SELECT row_count FROM __TABLES__, this &quot;job lifecycle&quot; introduces massive overhead. We were seeing latencies of <strong>~0.7 seconds</strong> for queries that should have been instant. The execution time was milliseconds, but the orchestration took the rest.</p><h3>The Solution: Short Query Mode (jobs.query)</h3><p>BigQuery offers a standard REST API endpoint jobs.query that supports a parameter jobCreationMode. By setting this to JOB_CREATION_OPTIONAL, you tell BigQuery: <em>&quot;If this query is short and fast, just run it and give me the results immediately. Don&#39;t create a Job ID unless necessary.&quot;</em></p><p>This bypasses the job creation and polling entirely for eligible queries.</p><h3>The Implementation</h3><p>At the time of writing, we couldn’t find a way to verify short query mode using the standard methods in the official PHP library. So, we rolled up our sleeves and implemented a lightweight wrapper using the Guzzle HTTP client to hit the REST API directly.</p><p>Here is the core of our BigQueryShortQuery class that made it possible:</p><pre>class BigQueryShortQuery {<br>    // ... setup and auth omitted for brevity ...<br><br>    public function query($sql, $options = [])<br>    {<br>        $url = &quot;https://bigquery.googleapis.com/bigquery/v2/projects/{$this-&gt;projectId}/queries&quot;;<br><br>        $body = [<br>            &#39;query&#39; =&gt; $sql,<br>            &#39;jobCreationMode&#39; =&gt; &#39;JOB_CREATION_OPTIONAL&#39;, // The magic switch<br>            &#39;useLegacySql&#39; =&gt; false,<br>            &#39;location&#39; =&gt; $options[&#39;location&#39;] ?? &#39;US&#39;,<br>        ];<br><br>        try {<br>            $response = $this-&gt;httpClient-&gt;post($url, [<br>                &#39;headers&#39; =&gt; [<br>                    &#39;Authorization&#39; =&gt; &#39;Bearer &#39; . $this-&gt;getAuthToken(), // Handles auth manually<br>                    &#39;Content-Type&#39; =&gt; &#39;application/json&#39;<br>                ],<br>                &#39;json&#39; =&gt; $body<br>            ]);<br><br>            $respBody = json_decode($response-&gt;getBody(), true);<br><br>            // If jobReference is missing, it meant the query ran instantly!<br>            if (isset($respBody[&#39;jobComplete&#39;]) &amp;&amp; empty($respBody[&#39;jobReference&#39;])) {<br>                // We normalize the compact JSON format (f/v) into a standard array<br>                return $this-&gt;normalizeRows($respBody[&#39;schema&#39;][&#39;fields&#39;] ?? [], $respBody[&#39;rows&#39;] ?? []);<br>            }<br><br>            // Fallback: If BigQuery decided to create a job (complex query), handle it as usual.<br>            if (isset($respBody[&#39;jobReference&#39;])) {<br>               // ... logic to handle long-running job if necessary<br>            }<br><br>        } catch (\GuzzleHttp\Exception\ClientException $e) {<br>            // Handle errors<br>        }<br>    }<br>}</pre><p>We completely bypass the heavy wrapper for these specific calls. We also had to implement a helper to normalize the raw JSON response (which comes in a compact f / v format) back into a friendly associative array, but it was a small price to pay.</p><h3>Results: 50% Lower Latency</h3><p>The impact was immediate.</p><p>For our dashboard counters and metadata checks, the average request time dropped significantly.</p><ul><li><strong>Before</strong>: ~700ms (Job creation overhead)</li><li><strong>After</strong>: ~150ms (Direct execution)</li></ul><p>That is a <strong>~78% improvement</strong> for those specific usage patterns, contributing to an overall application latency reduction of nearly <strong>50%</strong> for pages heavily relying on these metrics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*FqIiZZ4phVtXtWB1f7l5bw.png" /></figure><p>In the PHP world, where applications generate pages on-the-fly (“snap-on”), saving half a second on a heavy analytics page load is a game-changer for user experience.</p><h3>Conclusion</h3><p>If you are using BigQuery for user-facing applications or real-time dashboards, check if you are paying the “Job Tax” on small queries. Switching to Short Query Mode via the jobs.query endpoint might just be the &quot;free&quot; performance boost you were looking for.</p><p><em>The official library didn’t support it out of the box — I </em><a href="https://github.com/googleapis/google-cloud-php/issues/7604"><em>opened a request</em></a><em> for this feature to be implemented — but implementing the REST call was absolutely worth it.</em></p><h3>Wrap Up</h3><p>In the meantime, check it out other contents from me:</p><ul><li><a href="https://medium.com/p/195eb98c3d02">I Flipped a Switch in BigQuery and My Queries Got 30% Faster for Free</a></li><li><a href="https://medium.com/p/9e2634c84a82">Reduce your BigQuery bills with BI Engine capacity orchestration</a></li><li><a href="https://medium.com/p/def058a7c8dc">Hacking ADK’s Importer: How We Slashed 24-Second Cold Start in Half</a></li><li><a href="https://medium.com/p/cb03976f47c1">Introducing “GCP Project Switcher” Extension for VS Code — My Weekend Project</a></li><li><a href="https://medium.com/p/d14596ae28df">Announcing a New Datastore-Backed Session Service for the ADK</a></li><li><a href="https://medium.com/p/64da29fd58d1">Integrating DockerSlim container minify step on Cloud Build</a></li></ul><p>Feel free to reach out to me on <a href="https://www.linkedin.com/in/pentium10/">Linkedin</a> or read my previous posts on <a href="https://martonkodok.medium.com/">medium/@martonkodok</a></p><p><em>We at </em><a href="https://reea.net/"><em>REEA</em></a><em> are open to providing specialized knowledge, software development to your Google Cloud needs. Get in contact with Marton to kickstart discussions.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c29eee204b0a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/how-we-halved-latency-in-php-with-bigquery-short-query-mode-c29eee204b0a">How We Halved Latency in PHP with BigQuery Short Query Mode</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hacking ADK’s Importer: How We Slashed 24-Second Cold Start in Half]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://faun.pub/hacking-adks-importer-how-we-slashed-24-second-cold-start-in-half-def058a7c8dc?source=rss-350650c446a8------2"><img src="https://cdn-images-1.medium.com/max/1408/1*DYdyDCrQXNFI9b8eF8tR-Q.png" width="1408"></a></p><p class="medium-feed-snippet">What follows is an experimental and unofficial workaround for a specific use case: reducing cold start times for an ADK application on a&#x2026;</p><p class="medium-feed-link"><a href="https://faun.pub/hacking-adks-importer-how-we-slashed-24-second-cold-start-in-half-def058a7c8dc?source=rss-350650c446a8------2">Continue reading on FAUN.dev()  »</a></p></div>]]></description>
            <link>https://faun.pub/hacking-adks-importer-how-we-slashed-24-second-cold-start-in-half-def058a7c8dc?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/def058a7c8dc</guid>
            <category><![CDATA[adk]]></category>
            <category><![CDATA[google-adk]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[vertex-ai]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Fri, 31 Oct 2025 21:36:54 GMT</pubDate>
            <atom:updated>2026-01-25T21:14:17.859Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Announcing a New Datastore-Backed Session Service for the ADK]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/google-cloud/announcing-a-new-datastore-backed-session-service-for-the-adk-d14596ae28df?source=rss-350650c446a8------2"><img src="https://cdn-images-1.medium.com/max/1408/1*avYKhCAfcCJykQx8do200w.png" width="1408"></a></p><p class="medium-feed-snippet">I needed a way to maintain conversation history that was both persistent and cost-effective, without requiring a dedicated, 24/7 running&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/google-cloud/announcing-a-new-datastore-backed-session-service-for-the-adk-d14596ae28df?source=rss-350650c446a8------2">Continue reading on Google Cloud - Community »</a></p></div>]]></description>
            <link>https://medium.com/google-cloud/announcing-a-new-datastore-backed-session-service-for-the-adk-d14596ae28df?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/d14596ae28df</guid>
            <category><![CDATA[google-adk]]></category>
            <category><![CDATA[firestore]]></category>
            <category><![CDATA[cloud-firestore]]></category>
            <category><![CDATA[agent-development-kit]]></category>
            <category><![CDATA[google-cloud-datastore]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Sat, 27 Sep 2025 21:46:29 GMT</pubDate>
            <atom:updated>2025-09-29T02:21:43.748Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[I Flipped a Switch in BigQuery and My Queries Got 30% Faster for Free]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://martonkodok.medium.com/i-flipped-a-switch-in-bigquery-and-my-queries-got-75-faster-for-free-195eb98c3d02?source=rss-350650c446a8------2"><img src="https://cdn-images-1.medium.com/max/884/1*RKcUdZkB6RGlCm-miiwDJg.png" width="884"></a></p><p class="medium-feed-snippet">The Advanced Runtime is the next evolution of this powerhouse engine&#x200A;&#x2014;&#x200A;a complete modernization of its core technology</p><p class="medium-feed-link"><a href="https://martonkodok.medium.com/i-flipped-a-switch-in-bigquery-and-my-queries-got-75-faster-for-free-195eb98c3d02?source=rss-350650c446a8------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://martonkodok.medium.com/i-flipped-a-switch-in-bigquery-and-my-queries-got-75-faster-for-free-195eb98c3d02?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/195eb98c3d02</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[bi-engine]]></category>
            <category><![CDATA[query-optimization]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[bigquery]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Sat, 27 Sep 2025 11:48:23 GMT</pubDate>
            <atom:updated>2025-10-02T01:50:25.853Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Export BigQuery Saved Queries programmatically]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/google-cloud/export-bigquery-saved-queries-programmatically-ac0d7c28aa63?source=rss-350650c446a8------2"><img src="https://cdn-images-1.medium.com/max/1369/1*Bx4za9Wt9fkOZ-zk_PMz7w.png" width="1369"></a></p><p class="medium-feed-snippet">Hands on tutorial how to export/backup using the Dataform API your BigQuery saved queries.</p><p class="medium-feed-link"><a href="https://medium.com/google-cloud/export-bigquery-saved-queries-programmatically-ac0d7c28aa63?source=rss-350650c446a8------2">Continue reading on Google Cloud - Community »</a></p></div>]]></description>
            <link>https://medium.com/google-cloud/export-bigquery-saved-queries-programmatically-ac0d7c28aa63?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac0d7c28aa63</guid>
            <category><![CDATA[bigquery]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[google-cloud]]></category>
            <category><![CDATA[dataform]]></category>
            <category><![CDATA[data]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Mon, 27 Jan 2025 02:16:43 GMT</pubDate>
            <atom:updated>2025-09-30T05:44:42.934Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Imagen 3 generated images for restaurant/delivery apps: Eeatingh case study]]></title>
            <link>https://medium.com/google-cloud/imagen-3-generated-images-for-restaurant-delivery-apps-eeatingh-case-study-1c1ae98a955c?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c1ae98a955c</guid>
            <category><![CDATA[imagen]]></category>
            <category><![CDATA[vertex-ai]]></category>
            <category><![CDATA[gemini]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[google-cloud-platform]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Sat, 28 Sep 2024 05:10:29 GMT</pubDate>
            <atom:updated>2025-02-19T10:05:53.658Z</atom:updated>
            <content:encoded><![CDATA[<p>Product images can make the difference between a successful sale and traffic that materializes into nothing. But what happens when they are missing? With the help of AI technology, the <a href="https://eeatingh.ro/">Eeatingh</a> app now comes up with an innovative solution: automatically generating stylized product images, transforming simple, already-available data into eye-catching visuals that enhance the appeal of any menu. Let’s see how in the following case study.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1021/1*mQUPAPk4LtZW2FmwJ4nxkA.png" /></figure><blockquote>Credits: <a href="https://www.reea.net/">REEA.net</a> is a romanian software company, the author <a href="https://medium.com/u/350650c446a8">Márton Kodok</a> leads the Google Cloud platform engineering team, <a href="https://eeatingh.ro/">Eeatingh</a> platform made by REEA.</blockquote><h3>The challenge of missing product images</h3><p>The image is most often the first thing that grabs a potential customer’s attention. All the more true in restaurant apps, where users make decisions based largely on the visual appearance of dishes. Quality visuals play a key role in creating an attractive and compelling user experience, directly influencing sales. Without them, products appear incomplete, less appetizing, which can significantly reduce the chances of them being ordered.</p><p>Unfortunately, many restaurants struggle to provide images for all menu items. The main reasons are a lack of resources to produce professional photo shoots or the inability to quickly update images as the menu changes. The lack of these images not only affects the user experience, but can also create an impression of carelessness or unprofessionalism on the part of the restaurant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zZSMbwuRuFLfpQ_VS_SCSw.png" /><figcaption>Ciorba de fasoale boabe cu afumatura | Pizza vegana</figcaption></figure><h3>Meet the customer — the Eeatingh platform</h3><blockquote><a href="https://eeatingh.ro/">Eeatingh</a> — by REEA, is a romanian local online platform that has been delivering your favorite food straight to customers’ doorsteps since 2014. In its 10 years of existence, Eeatingh has successfully competed in the local market with big players such as Food Panda, Glovo and Tazz, maintaining its position through an agile and flexible business model.</blockquote><blockquote><strong>Despite the presence of these giants,</strong> <strong>Eeatingh has thrived by offering a digitized home delivery service at no extra cost to the end customer, allowing users to enjoy delicious meals without leaving the comfort of home.</strong></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RPNQRZd4jUHUVRrqQM_yrg.png" /></figure><h3>Our solution: automatic AI image generation</h3><p>To solve the problem of missing product images, we developed an innovative solution that uses artificial intelligence to automatically generate stylized images based on product data in the Eeatingh app. Through <strong>an API built in Python</strong>, our app now utilizes a generative text-image AI model that transforms simple product information (such as name and description) into attractive images in a drawn or vector style.</p><p>The icing on the cake: the process doesn’t require manual intervention to create detailed prompts, as the API automatically picks up the product data and sends it to the AI model. Thus, restaurants that don’t have images for their entrees can benefit from a fast and efficient solution integrated directly into the Eeatingh platform, without any hours of prompt-writing training. Instead, it’s all done with a click of a button: within the admin section, there’s a new button that allows images to be automatically generated, offering a practical alternative to uploading photos manually.</p><p>The technology behind this solution includes <strong>Google Cloud — Vertex AI</strong> and integration with the <a href="https://deepmind.google/technologies/imagen-3/"><strong>Imagen 3</strong></a><strong> image generation model</strong>. This model utilizes an intermediate step with a Large Language Model (LLM), which expands short product names into detailed descriptions, including relevant information such as ingredients, how to prepare or serving suggestions. This data expansion significantly improves the quality of the generated images.</p><blockquote>As an aside, it is worth mentioning that <a href="https://deepmind.google/technologies/imagen-3/">Imagen 3</a> was recently launched in August 2024, and Reea was invited as a partner to test this technology, providing feedback and receiving suggestions and best practice models in its use from Google engineers.</blockquote><p>Through this technology, restaurants not only save time and resources, but are also able to provide users with an enhanced visual experience, thus helping to increase the attractiveness and conversion rate of their products.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XbYLLV_dhZCw8vO1DClMGQ.png" /><figcaption>Romanian mici | Turkish kebab</figcaption></figure><h3>How does the image generation API work</h3><p>The process starts with submitting product data — product name and description — to the API. This data is then processed by a large language model (LLM), which extends and enriches the initial information. For example, a simple product name such as “margherita pizza” can be transformed into a detailed description that includes information about the ingredients (mozzarella cheese, tomato sauce, etc.), the cooking method (baked on the stove) and the serving style (with fresh basil). This step is essential to generate the most realistic and relevant images.</p><p>Once the description has been generated and optimized, the API sends this data to the Imagen 3 model, which creates the product image. Imagen 3 is known for its ability to generate high quality images with a high level of detail and realism and is optimized to deliver fast and accurate results.</p><p>Our API can be used by multiple applications or websites and is designed to be flexible and easy to integrate. The associated costs are low due to the use of cloud services and optimizations in place, and for a high volume of images, the price per image can drop considerably, making this solution affordable for businesses of all sizes.</p><p>This entire image generation process takes place in the background, without requiring user intervention in the administration section of the Eeatingh application. Thus, when a product does not have an image uploaded, a new image can be automatically generated with a simple click, providing a fast and efficient solution for filling out restaurant visual menus.</p><h3>Challenges encountered and solutions found</h3><p>While automatic AI image generation brings significant benefits, the process of developing and integrating the technology was not without its challenges. Here are the biggest ones:</p><h4>1. The challenge of short or traditional product names</h4><p>One of the main challenges encountered was related to products with short or very generic names, such as “pizza” or “pasta”. These simple names did not provide enough context to generate a clear and detailed image, leading to average or irrelevant images. Another problem arose with traditional dishes with regional names, such as “tochitură” or “mămăligă”, which did not translate efficiently or were not correctly recognized by the AI models.</p><p>Solution: we implemented a large language model (LLM) that extends and enriches short names by automatically generating detailed descriptions. This naming expansion process includes specific ingredients, preparation methods, and serving suggestions, providing the AI model with enough information to generate more accurate and relevant images. Also, for traditional products, prompt enhancement mechanisms have been added, which can include cultural descriptions or examples of similar dishes from other cultures.</p><h4>2. Text on generated images</h4><p>Occasionally, some generated images contained superimposed text, which was undesirable, especially in the context of menu images for restaurants. It must be said that this is a specific problem encountered in generative models that are trained on data that also includes textual elements.</p><p>Solution: We added a quality control filter to the generated images, which detects and removes images with textual overlaps. Users are also provided with a feedback mechanism where they can flag inappropriate images, helping to continuously improve the generation and filtering process.</p><h4>3. Managing images for products with very similar names</h4><p>In some cases, products with very similar names (such as “pizza margherita” and “pizza quattro formaggi”) resulted in the generation of almost identical images, even though the products in question have significant visual differences.</p><p>Solution: we implemented an algorithm that takes into account key differences between products when expanding textual descriptions, ensuring that the AI model generates distinct and relevant images for each product. By including details such as specific types of cheese or different cooking methods, we were able to diversify the generated images even for similar products.</p><h4>4. Cost and real-time performance</h4><p>Generating AI images can be expensive, especially for large volumes of products. The speed of image generation is also critical, as restaurants need to be able to quickly add new dishes to menus without significant delays.</p><p>Solution: We opted for Google Cloud integration and the Imagen 3 model, which offers the optimal balance between quality and speed. The Imagen 3 Fast allows images to be generated in less than 5 seconds for a set of 4 images, which is ideal for restaurant needs. In addition, utilizing cloud services allows us to keep costs under control, ensuring a competitive price per image generated, with costs reducing as volume increases.</p><p>By solving these challenges, we were able to optimize the image generation process and provide an efficient, scalable and affordable solution for restaurants using the Eeatingh app.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mmZP5WuP752W8d8_1aS0OA.png" /></figure><h3>Use cases and long-term potential</h3><p>The automated AI image generation technology we’ve implemented in the Eeatingh app offers obvious immediate benefits, but its true potential is far greater. Here are a few examples:</p><ol><li><strong>E-commerce and retail</strong><br>One area with huge potential is e-commerce. Online stores, especially those with a large inventory of products, often struggle to present quality images of every item in stock. AI’s automated image generation solution can be used to quickly create attractive images for products that don’t have photos available. Whether it’s fashion items, technology, or even handcrafted products, this technology can turn textual descriptions into visual images that attract customers.</li><li><strong>Hospitality and tourism</strong><br>The hospitality and travel industries can also capitalize on automated AI image generation. Hotels can create images for rooms, facilities or special service packages based on their descriptions. Travel agencies can also generate images for different vacation packages or destinations that do not have real or updated photos, providing a visual representation to attract tourists.</li><li><strong>Online education and training</strong><br>Online education and training platforms can use this technology to create compelling visuals for courses, tutorials or educational materials that do not have images by default. For example, a course on cooking can utilize AI-generated images for each recipe or ingredient, giving learners a more engaging learning experience.</li></ol><h3>Long-term potential</h3><p>In the long term, AI image generation technology can become a standard tool for any business that relies on visual product presentation. Continuing developments in AI models, such as those offered by Google Cloud — Imagen 3, will enable the generation of increasingly realistic and detailed images at lower costs and faster speeds. Also, integration with platforms such as social media or marketplaces will enable the automatic generation of images for posts, advertisements and product pages, eliminating the need for external visual assets.</p><blockquote>Besides, at <strong>Reea </strong>we have already experimented with other tools, including in commercial projects where we have used AI models such as <strong>DallE</strong>, <strong>Midjourney </strong>and <strong>Stable Diffusion 3</strong> in logo design, marketing content, social media posts, etc. already demonstrating the potential of these technologies in multiple domains.</blockquote><p>Thus, from restaurants and online stores to hotels and educational platforms, automated AI image generation will continue to expand its applicability, becoming an integral part of many industries’ marketing and sales strategies.</p><h3>Wrap Up</h3><p>In the meantime, if you want to check it out, here are some links:</p><ul><li><a href="https://cloud.google.com/blog/products/ai-machine-learning/a-developers-guide-to-imagen-3-on-vertex-ai">A developer’s guide to getting started with Imagen 3 on Vertex AI</a></li><li><a href="https://cloud.google.com/vertex-ai/generative-ai/docs/image/overview">Imagen on Vertex AI</a></li></ul><p>Feel free to reach out to me on Twitter <a href="https://twitter.com/martonkodok">@martonkodok</a> or read my previous posts on <a href="https://martonkodok.medium.com/">medium/@martonkodok</a></p><p><em>We at </em><a href="https://reea.net/"><em>REEA</em></a><em> are open to providing specialized knowledge, software development to your Google Cloud needs. Get in contact with Marton to kickstart discussions.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c1ae98a955c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/google-cloud/imagen-3-generated-images-for-restaurant-delivery-apps-eeatingh-case-study-1c1ae98a955c">Imagen 3 generated images for restaurant/delivery apps: Eeatingh case study</a> was originally published in <a href="https://medium.com/google-cloud">Google Cloud - Community</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Automating table backups in VertexAI training tasks — through BigQuery]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/google-cloud/automating-table-backups-in-vertexai-training-tasks-through-bigquery-566c5d7c6da5?source=rss-350650c446a8------2"><img src="https://cdn-images-1.medium.com/max/1371/1*ZQQomutPaJ72VxVZimNyFA.png" width="1371"></a></p><p class="medium-feed-snippet">Discover the advantages and the necessity of backing up a table during its usage in a VertexAI training task. We at REEA implemented an&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/google-cloud/automating-table-backups-in-vertexai-training-tasks-through-bigquery-566c5d7c6da5?source=rss-350650c446a8------2">Continue reading on Google Cloud - Community »</a></p></div>]]></description>
            <link>https://medium.com/google-cloud/automating-table-backups-in-vertexai-training-tasks-through-bigquery-566c5d7c6da5?source=rss-350650c446a8------2</link>
            <guid isPermaLink="false">https://medium.com/p/566c5d7c6da5</guid>
            <category><![CDATA[google-cloud-platform]]></category>
            <category><![CDATA[bigquery]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[cloud-workflow]]></category>
            <category><![CDATA[vertex-ai]]></category>
            <dc:creator><![CDATA[Márton Kodok]]></dc:creator>
            <pubDate>Sun, 26 Feb 2023 22:09:09 GMT</pubDate>
            <atom:updated>2023-02-28T09:21:10.821Z</atom:updated>
        </item>
    </channel>
</rss>