Wednesday, February 28, 2024

Driving Efficiency with LLMs: Transforming Complex Content into Digestible Summaries

Hey everyone,

I recently shared my latest side project with you all here . It's a simple web page where I've been experimenting with various tools, and my latest addition focuses on AI. Lately, I've been diving deep into the realms of AI, machine learning, and data science, exploring the tools and libraries that are making waves in the tech world. Like many, I'm captivated by the possibilities offered by transformers and LLMs, and I'm eager to integrate them into real-world applications to tackle previously unsolvable problems. We have new tools now, let's see what has changed.

Of course, mastering these technologies doesn't happen overnight. I'm dedicating time to different projects, using them as opportunities to learn, experiment, and refine my skills. In my last blog post , I discussed automating invoice entry, or perhaps "assisting" is a better term. The goal was to extract values from documents and streamline data entry, ultimately making life a little easier and saving time.

This time around, I'm delving into LLMs with a focus on summarization. I've developed and deployed a tool capable of summarizing content from YouTube videos, PDF files, or any web page. The concept is straightforward: extract the text and condense it down. To streamline the process, I'm leveraging 🦜️🔗 LangChain under the hood, which simplifies the task. Here's a glimpse into some of the tools and APIs I've utilized:

1 For YouTube transcript extraction, I've tapped into the youtube-transcript-api , allowing users to download transcripts if available. Transcript can be uploaded by the user or auto-generated, and it can be in multiple languages, and it can be translated to other langueages as well. All this can happen on the fly. Transcribing the auido/video yourself is also an option, there are a lot of models trained for this purpose paid like Oracle AI Speech and open source like Whisper from OpenAI, and many more...

2 Processing PDF documents has been an enjoyable challenge. I was already experimenting extensively with PyMuPDF and now PyPDFLoader . One thing I like about python world is there are hundreds of libraries out there, you can find a huge list of document loaders here .

3 Web page processing required leveraging AsyncHtmlLoader and Html2TextTransformer and beautiful soup to sift through the HTML and focus solely on the textual content.

4 Summarization duties are handled by LLMs (such as OpenAI and Cohere), albeit with a limited context window. If the text fits into LLM context it is great, if not then to work around this, I've implemented a technique called "MapReduce ," chunking the text into manageable pieces for summarization, then condensing those summaries further. This might not be required at all in the near future, competition is tough and everyday we witness increasing model parameters and context windows .

Image taken from

5 I've added a few user-friendly features, such as automatically fetching and displaying thumbnails for pasted links. Whether it's a YouTube video, PDF file, or web page, users will see a visual preview to enhance their experience. YouTube offers thumnails if you know the video_id, which part of of the URL. For PDF files I've created a thumbnail image of the first page. Web pages were a bit tricky, I used headless chromium to get screenshot of the page. This Dockerfile was extremely helpful to achieve this.

6 To keep users engaged while awaiting results, I've enabled streaming via web sockets using for both console logs and chain responses. Although it does come with a cost, it is highly motivating for user to keep waiting as witnessing things happen under the hood. And with ChatGPT, it kind of became the defacto.

7 And finally everything is nicely packed in a container and deployed on Oracle Cloud behind load balancers secured with TLS/SSL and CloudFlare etc. You can read about the setup here .

As always, I've shared all the code on my GitHub repository , along with the references that aided me at the bottom of this post. I'll also be putting together a quick demo to showcase these features in action. While these concepts may seem generic, practice truly does refine them. Implementing these tools has not only boosted my confidence but also honed my skills for future challanges.


I hope this project serves as inspiration for readers, sparking ideas for solving their own challenges. Perhaps it could fit into any requirement that needs simplifying the review and categorization of lengthy text. Documents like Market Research Reports, Legal Documents, Financial Reports, Training and Educational Materials, Meeting Minutes and Transcripts, Content Curation for your Social Media and even Competitive Intelligence Gathering...

All we need is a little creativity, and courage to tackle our old unsolvable problems with our new tools. Langchain itself is using this to improve documentation quality combined with clustering/classification logic.

If you have any ideas or questions, feel free to discuss them in the comments. Don't hesitate to reach out; together, we might just find the solution you're looking for.

Happy coding!

1.Developing Apps with GPT-4 and ChatGPT Most of the foundation and ideas came from this book, highly recommended for beginners. It is available on O'Reilly platform
2.🦜️🔗 LangChain You will find almost all you need, getting started guide, sample codes for tools, agents, llms, loaders ...
3.PyMuPDF PDF Library
5.Document Loaders
6.html2image for web page thumbnails Web Sockets
8.I, Robot for PDF testing
9.Streaming for LangChain Agents Video Tutorial by James Briggs
10.YouTube Transcript API

Monday, February 12, 2024

APEX with AI Services: Automating invoice entry with AI assisted key/value extraction

Since couple of years, almost everyone is talking about AI, trying to understand how this can help, both life and business. OpenAI ChatGPT made a huge impact in our lives, "You are a helpful assistant", thank you! I am using it daily. I have been also playing around with other language models like Llama 2 and actively learning ML from different resources. Hugging Face is a must join platform along with all the material on LangChain . Just these two got me as far as I could train and run my own model to classify my emails within a week, and the results were incredibly better than I could ever expect. Besides I am enjoying this a lot.

Nowadays the number of interested customers is increasing and this post is about a very basic customer use-case, a real one: invoice entry. I know, it doesn't sound interesting, at first I thought my technical consultancy for ERP days are over, but I promise this is not boring. It has new challanges for me (and for most customers) and demonstrates application of AI services to real life problems. So let's dive into it!


"...investigating the possibilities of automating / optimizing the reading and processing of PDF documents with the help of Optical Character Recognition (OCR)..."
The moment I saw this I could imagine what they wanted. After verifying their ideal soution in our discovery meeting, we planned for a demo to proove it can work.
Here is a mock design, on the right side we display the PDF file, OCR'ed and all values extracted, and on the left side form is populated with the extracted values. Ideally operator will just click save, and will have a chance to fill in any missing information, huge time saving.


Biggest challange is lack of skills, customer knows APEX inside out but I am not an APEX developer. I understand the APEX environment, components and how they work, installed and configured many times. Followed and demonstrated many workshops, but never developed something from scratch. Yet APEX is low code, there are many samples and I was able to complete this in 2 days.


I will briefly mention the steps I've followed and highlight the important parts. Using cloud services makes it easy to start.
1 I start with creating an Autonomous database and APEX workspace. It takes minutes to start worrking on my APEX development.
2 Then I follow this LiveLab workshop as a starter application. I tweaked the table structure according to my needs, but it gave me the foundation I needed for interacting OCI Document Understanding service using object storage is a good decision.
3 Using Document Understanding service inside APEX was easy.
I have added API endpoint in application definition.

Then I made changes to saveDocument process to invoke the AI service after uploading file to Object Storage. The below code prepares the JSON body, invokes the service processorJobs, parses return message and updates the table. The execution flow and response time is not efficient for production but good enough for a POC. I am using only Key Value Extraction feature, but you can also use all features including generation of a searchable PDF file in your output bucket. The service is pretrained and capable of identifying common key/value pairs from an invoice document. My service call creates a json file, which can be located with the job_id. I've also loaded that json file into a blob, parsed and created views on top of it just to make my life a bit easier. Chris Saxon has an excellent cheat sheat for that purpose.

4 I created a new page, with 3 regions. Two side by side, left for showing/entering extracted values, right an iframe to display PDF file, last for invoice lines, as designed in the mock wireframe.
For displaying PDF inline on the right side of the page I followed instructions in this YouTube video . The only thing is I didn't have a link item on the same page, but the ID has to come as a page parameter. So I added the link on my home page where all uploaded files are listed, and passed the document_id as page parameter. Then created a new Page Load Dynamic Action to get ID and trigger PDFViewer Action to display the file.
After changing the theme to Redwood (with some modifications to make it dark) the application looks like this:


APEX and AI Services is a very powerful combination, that can help you boost the productivity. Please share your ideas about what you think and of course new use-cases in the comments, maybe we can build one together!

1. LiveLabs: Use OCI Object Storage to Store Files in Oracle APEX Applications
2. OCI Developer Documentation: Document Understanding API
3. OCI Developer Documentation: Key Value Extraction (Invoices)
4. Oracle Blog: How to Store, Query, and Create JSON Documents in Oracle Database
5. Oracle Developer YouTube Channel: PDF Viewer in Oracle APEX
6. Banner Image Credit: Dall-e 3
7. YouTube Background Soundtrack: BenSound: Royalty Free Music for Videos
8. Postman Collections: OCI REST API Postman Collection

Saturday, February 3, 2024

From Mock API Server to Mandelbrot Sets over the Load Balancer, CloudFlare, Let's Encrypt and Beyond...

Hello everyone! It's been a while since I last shared my thoughts, but I've been deep in the trenches of a fascinating new stuff. You know how things are, it started with a simple need for a mock API server and led me through setting up load balancers, domain management, and fun with Mandelbrot sets. Let me take you through it. This can be used as a blueprint for deploying your web workload securely on modern cloud environments.

Setting the Stage with a Python Flask App

Normally a quick dirty python http.server is enough for testing accessibility but this one time we needed a bit more than that. It all began with a customer request that they needed to setup a mock API server. A Python Flask app seemed to be a suitable option for that time: I've never used it before. I must admit curiosity and growing pythonic love inside me made me do this choice. This initial step, though seemingly straightforward, laid the groundwork for the exciting challenges that followed. There are lots of tutorials to get you started, just start with this link.

Load Balancers, Domains, and the Birth of CodeHarmony.Net

I was just playing around with my new Flask app, discovering blueprints , jinja templates etc., you know usual things. Then another customer came with a fair request:

I want to use my load balancer with multiple backends accessible via different domains and subdomains.
Which is possible with Virtual Hostnames . Start with adding your hostnames (actual DNS A Records).
Then you need to decide if
  • you want to use one listener for each backendset (which implies each listener will be using a different port) e.g.,, etc... Straightforward if you are okay with different ports. Just select the hostname(s) during listener creation.
  • you want to use one listener for all backendsets (which implies you should be able to distinguish by the requested URI). This one a little tricky and requires proper planning. You need to combine hostnames and path route sets. Here is the official documentation
For testing this in my tenancy I needed a domain, well, mmm... Maybe not, I could just edit my host file but to be honest I wanted to also test deploying my Flask app and see it live instead of scraping it. This prompted the creation of CodeHarmony.Net . I followed the hype, and the name born out of collaboration with ChatGPT. With a smooth purchase from Namecheap , CodeHarmony became the hub for my tech experiments

Navigating the Cloud with Cloudflare and SSL/TLS Encryption

You know, these are the basic steps to ensure a secure and efficient web deployment.

  • I created a free Cloudflare account, with basic protection to proxy connections to my loadbalancer as I don't want them to be exposed to all internet directly.
  • Then I updated nameservers on my domain name provider (Namecheap) with the ones CloudFlare provided.
  • Then created domain and subdomains (A Records) using my Load Balancer Public IP
  • Then proxied the connection to load balancer. I implemented full strict SSL/TLS encryption, terminating at the load balancer (which means private subnet traffic is not encrypted)
  • On load balancer end, configured listeners for both HTTP and HTTPS. I mentioned a little about routing above.
  • To make sure all traffic is supported and encrypted I also created an HTTP listener and redirected all HTTP to HTTPS using URL Redirect Rules
  • Cloudflare brings it own TLS certificates which is very cool. For my loadbalancer HTTPS Listener I used a free Let's Encrypt certificate.
  • I restricted access to my loadbalancer to CloudFlare IP addresses (I am not sure if I whitelisted all of them though...). One cool thing I noticed while investigating loadbalancer access logs, CloudFlare sends end client IP address in forwardedForAddr field. Here is a simplified access log:

Flask Web App Deployed on Container Instances as Backend

For those who share my passion for coding adventures, all code is available on my Github repo . Dive in, explore, and feel free to contribute. Dockerfile is also included as it is packaged as a container. You will also find a file which is building the container image, pushing it to a private OCI registry then creating container instance. You need to prepare .env file which is not included in the repo, the file should contain following values: Once set you should be able to build and deploy your application. Just notice that I am using London region, if you use another remember to change values including OCI Registry.

It was super fun to design, code, build and deploy everything with the best practices applied. It felt good when put myself to test and find out that I can still code Full Stack with a basic Python/Flask backend with Bootstrap and JQuery for the frontend. I was a bit rusty in the begining but progressed very fast. Admittedly, React or Vue might have been the trendy choices, but the experience proved invaluable, reigniting my confidence in full-stack development.

You will find many scripts converted into tools in these pages: manipulating PDF files, playing around with images. Captivating world of Mandelbrot sets which is a beauty of drawing, painting, and exploring color palettes became my newfound fascination, leading to the discovery of hidden gems waiting to be unveiled.

Curious about the technical details? Check out the final topology:

Until next time, keep coding, stay curious, and embrace the ever-evolving tech landscape.

Happy coding! 🚀

Thursday, November 2, 2023

OCIFS: Using Object Storage as local file system

For using object storage like a file system, we used to and still have Storage Gateway to mount OCI Object Storage like an NFS mount target. It was only available on Linux 7, it required Docker and some resources (not small). For Windows rclone , a third party solution was the answer, and I wrote about it here in one of my old blog post

Now we have another alternative, recently announced OCIFS, pretty much does the same thing. It requires Oracle Linux 8 or later, and less demanding resources.

1 Installation is easy, on OLE8 instance, I ran dnf install command. And as you see it has a very small size.

[opc@ocifs-demo ~]$ sudo dnf install ocifs
Total download size: 156 k
Installed size: 360 k
Is this ok [y/N]: y
Downloading Packages:
(1/2): ocifs-1.1.0-2.el8.x86_64.rpm               824 kB/s |  73 kB     00:00
(2/2): fuse-2.9.7-16.0.1.el8.x86_64.rpm           446 kB/s |  83 kB     00:00
Total                                             824 kB/s | 156 kB     00:00

2 For authentication I used API Key method. I have just copied my OCI CLI config file and key file to default locations, so I didn't have to pass any parameters. I didn't even install cli itself. Then I mount object storage bucket "ocifs-mounted-bucket" to "mydir" folder

[opc@ocifs-demo ~]$ ocifs ocifs-mounted-bucket mydir
[opc@ocifs-demo ~]$ cd mydir/
[opc@ocifs-demo mydir]$ mkdir new-folder-01
[opc@ocifs-demo mydir]$ mkdir new-folder-02
[opc@ocifs-demo mydir]$ history > new-folder-01/history.txt


If your instance is running on OCI, you can also use instance principals for authentication

ocifs --auth=instance_principal ocifs-mounted-bucket mydir

For non-default config file, you can use config parameter for passing config file location

ocifs --auth=api_key --config=~/my_config ocifs-mounted-bucket mydir

3 Can be unmounted with any of these commands

fusermount -u mydir
sudo umount mydir

By the way, there is a Python library with the same name, ocifs which also enables python to use object storage like a file system.

1. OCI Documentation: Storage Gateway
2. RClone: Mounting Object Storage on Windows
3. OCI Blog: Introducing OCIFS
4. OCI Documentation: OCIFS Utility
5. OCI Documentation: Install OCI CLI on OLE8
6. Oracle GitHub: OCIFS Python Library

Wednesday, October 11, 2023

How fast can I launch multiple OCI compute instances using Java SDK? #JoelKallmanDay

When I saw Tim Hall's blog post about #JoelKallmanDay it touched my heart. If you are somehow interested with Oracle APEX you probably know who Joel Kallman is. He meant a lot to the community. He is missed by the people all around the world, by the people he didn’t meet face to face. I wish this blog post was about APEX, maybe next year...

This one was waiting in my stash for long time as I wasn't happy about the dirty POC code and didn't have the time to refactor it. Actually it is about a really niche and cool requirement and second part of something I've posted in the past .

So let me start with a little context. Can you imagine how the big e-commerce platforms get ready for their peak seasons? This is about a software house who is highly specialized in load testing e-commerce applications. They have their own platform where e-commerce users prepare their test scenarios, and launch hundreds of thousands individual web agents to test the application for like 30 minutes. Under the hood the actual testing platform provisions tens (or hundreds) of compute instances, deploy the test code and runs it. Once the desired testing duration ends, all the compute instances are terminated. Perfect use-case that's possible only on cloud! This was one of the cool use-cases I've seen. Although the application is polyglot, they have chosen Java to code the instance creation part. So here we start.

My starting point is as usual OCI Online Documentation SDK for Java. The documents have links to Maven repository where I can just include the dependencies in my POM file . And there is an Oracle GitHub repository with quick start, installation and examples that will get me started in minutes. I've quickly located example code for creating a compute instance. The sample code is huge, it is creating everything from scratch, not only the compute instance but also VCN, subnet, gateways, etc. It is a comprehensive example, kudos to the team.

1I need a very quick test on how fast I can create instances. So here is a simplified test code which is getting all required inputs from environment variables (region, AD, subnet, compartment, image, and shape identifiers already that already exist), original sample is using waiters so I keep it just to see how convenient to wait my instances to reach a certain state (running)

And if I just test it with 5 instances to be created, the output is:

----------------------------------------------------------------------- created in 36244 ms created in 32845 ms created in 32102 ms created in 31995 ms created in 62075 ms
Total execution time in seconds: 196

I am provisioning instances one by one and waiting for the instance to transition into RUNNING state. It took around ~30 seconds to provision a compute instance and see it in running state. Not bad at all. But this is not good enough, for extreme cases my customer needs tens of instances, can we do better?

2So I think I don't need to wait for the compute instance to reach running state before provisioning the other one, as long as I have the OCIDs of instances, I can come back to check the state later.

This time since expecting to wait less, I test it with 10 instances. Here is the output:

----------------------------------------------------------------------- created in 2427 ms created in 878 ms created in 1041 ms created in 982 ms created in 971 ms created in 772 ms created in 743 ms created in 754 ms created in 972 ms created in 812 ms
Total execution time in seconds: 12

This is a lot better, it is down to ~1 second per instance from 30 seconds per instance. I wonder if this can get any better. It is still synchronous call, one by one.

3What happens if we make it asynchronous? For this purpose I am using AsyncHandler which enables you with callback functions. Compute client also takes a different form: ComputeAsyncClient, input is the same. I do some concurrent processing with Futures , just to see if threads are done and collect the compute instance OCIDs

I again test it with 10 instances. Here is the output:

work requested in 391 ms
work requested in 14 ms
work requested in 10 ms
work requested in 9 ms
work requested in 7 ms
work requested in 7 ms
work requested in 5 ms
work requested in 6 ms
work requested in 7 ms
work requested in 4 ms
test-9 -
test-10 -
test-1 -
test-4 -
test-5 -
test-6 -
test-7 -
test-8 -
test-3 -
test-2 -
Total execution time in seconds: 2

As you can see from the output, there is no order because it is asynchronous and randomly created depending on thread execution order. It is blazing fast, took 2 seconds in total to create 10 instances!


What if I get greedy and try a larger batch? Then I get an error message because of request throttling protection.

Here is a little script to clean-up that can be used during tests.

1. OCI Documentation: SDK for Java
2. Oracle GitHub Repository: SDK for Java
3. Oracle GitHub Repository:
4. Tutorial: java.util.concurrent.Future
5. OCI Documentation: Request Throttling
6. OCI Documentation: Finding Instances

Monday, October 9, 2023

How to clone boot volume cross tenancy including Free Tier

In this blog I try to write about unusual things, not always possible though. I prefer to write beacuse mostly for myself to remember what was the solution, second to share with friends and customers. This is one of the interesting ones.

The question is "One of my ex-employees has a demo environment in his Free Tier tenancy (which means seeded credits already spent/expired) and I want to move the compute instance (Always Free Micro Shape) to my paid company tenancy". If you take a close look at the documents you will find out that the block volume can be replicated accross data centers and regions . Volume backups are regional but you might also copy accross regions . But this is only possible within the tenancy. To be honest, this is strange because some customers use OCI Cloud with Organizations , parent/child relationship of their tenancies. But Free Tier is a blocker.

Next thing that comes to my mind is creating a custom image, and export/import image using Object Storage as explained here .

But as you see, since it's a Free Tier tenancy now we don't have the limit and the motivation.

So while searching for alternative, talking to PM I came accross this undocumented feature . Basically the solution playbook is saying if you setup proper policies in both tenancy (define the other tenancy and authorize it to access the resources), then using the cli or API you can clone a volume from one tenancy to another. Or restore a volume backup from tenancy to the other. So here is what I did.

1I have created the following policy in source Free Tier tenancy, the policy defines the target tenancy and authorize a group in target tenancy to clone a volume

Define tenancy NewTenancy as $TARGET_TENANCY_OCID
Define group NewTenancyIdentityGroup as $TARGET_TENANCY_GROUP_OCID
Admit group NewTenancyIdentityGroup of tenancy NewTenancy to use volumes 
in tenancy where ANY { request.operation='CreateVolume', 
request.operation='GetVolume', request.operation='CreateBootVolume', 
request.operation='GetBootVolume' }

2I have created the following policy in target tenancy, the policy defines the source tenancy and authorize a group to clone a volume, very similar to first one.

Define tenancy OldTenancy as $SOURCE_TENANCY_OCID
Endorse group NewTenancyIdentityGroup to use volumes 
in tenancy where ANY { request.operation='CreateVolume', 
request.operation='GetVolume', request.operation='CreateBootVolume', 
request.operation='GetBootVolume' }

3Then invoked the API with CLI to clone the boot volume in source tenancy ($BOOT_VOLUME_ID) with my profile connected to target tenancy

oci bv boot-volume create --profile=cross_tenancy_user_profile --debug \
--region=eu-frankfurt-1 --source-boot-volume-id $BOOT_VOLUME_ID  \
--display-name Cross-Tenancy-vm-e2micro-5 --compartment-id $COMPARTMENT_ID

1. Don't forget using compartments for your cli command
2. Also make sure the Group you are using in your target tenancy and the profile user can create block volume
3. If you get 404 - NotAuthorizedOrNotFound error message, most likely related to your policies
4. Policies are replicated to other regions from Home region, if you are working on a different region than your home region, take that into consideration
5. For same AD use clone, for different AD use backup restore
6. Although this seems to be the only way to copy block volume from a Free Tier without converting it to a paid tenancy, this feature can be very useful for moving large boot volumes and for bulk operations to move multiple boot volumes. It will be definitely easier than using Object Storage to export/import images which has a size limitation also.
7. Just imagine what other interesting use cases can be achieved with this admit/endorse policy setup

1. Solution Playbook: Migrate Oracle Cloud Infrastructure volume data across tenancies
2. OCI CLI Command Reference : boot-volume » create
3. OCI Block Volume Documentation: BYOI Best Practices
4. OCI Block Volume Documentation: Copy Block Volume

Wednesday, October 4, 2023

Should I use security list or network security group or both to secure my OCI deployment?

Today I was in a customer call, it was a pretty straightforward scenario. The session turned into hands-on pretty fast, and I love it, sharing the curiosity and eagerness to solve the problem with technical people, few things can match that feeling. And as always we came to the point where we delve into the troubleshooting. So here we go...

The requirement is simple, deploy an Ubuntu server to host a demo application over HTTP port 80, a small VM in a public subnet with a public IP Address and supporting security rules. VCN is created with the wizard, and it comes with a Default Security List which is populated with 3 stateful ingress rules:

First rule enables SSH access to my host, the other two ICMP rules are there for debugging and they don't enable a ping response. All of them are stateful. And this is the Egress part:

There is one stateful egress rule which enables outgoing traffic to any destination with any protocol on any port. State will be important as we will find out later...

Security list is attached to subnet and enforced at all VNICs in the subnet. So setting general rules with security list makes sense, however we also need to open HTTP 80 port for one server and we don't want this for all servers in the subnet. For this purpose we use network security groups which is another type of virtual firewall that Oracle recommends over security lists. You can use security lists and network security groups together. How do the rules apply? At simplest: a union of all rules are applied to VNIC. Security list is tied to subnet so applies to all VNICs in the subnet, NSG is attached to individual VNIC, so it's granular. Here are the rules in our NSG:

First rule is allowing incoming TCP traffic on port 80 from any source, the second rule is allowing outgoing TCP traffic. And rules are stateless , which means connection tracking is disabled. Why would I want that? Maybe I am expecting high traffic, or maybe I was greedy and wanted everything at once.

Overall architecture can be simplified like this:

So we SSH into our Ubuntu server using our public IP, also add linux firewall rules by updating iptables as explained in detail on this tutorial .

All set, for a really quick dirty test, let's run python

And it works, but we quickly find out there is another problem. We can't access Ubuntu repositories to update the packages or install new ones. Although IPv6 in the error message is distracting, it doesn't work with IPv4 either. It is a problem with accessing the internet.

So after some debugging, we soon realize the problem is having overlapping stateful and stateless rules. Our stateful egress rule on the security list should be providing all the access we need towards internet. But it doesn't, why? Because our stateless egress rule in NSG is overlapping and overriding the SL as stateless has precedence over stateful. This is what documentation exactly warns us about.

If for some reason you use both stateful and stateless rules, 
and there's traffic that matches both a stateful and stateless rule 
in a particular direction (for example, ingress), the stateless rule 
takes precedence and the connection is not tracked. You would need 
a corresponding rule in the other direction (for example, egress, 
either stateless or stateful) for the response traffic to be allowed.

Lessons learned
1. Use stateful rules (which is default) unless I have a good reason to use stateless
2. If using stateless ingress always exactly match it with an egress rule, don't use a broader rule
3. Don't use overlapping stateless and stateful rules, as the stateless rule takes precedence and the connection is not tracked, thus acting different than expected.

How did we fix it?
On our NSG, we converted ingress rule from stateless to stateful, and removed egress rule as it's not needed anymore.

If we wanted to use stateless rules, then a viable solution will be restricting egress rule to exactly match the ingress thus protocol TCP, source port 80 and destination port any.

1. OCI Security Rules: Stateful Versus Stateless Rules
2. Developer Tutorials: Free Tier: Install Apache and PHP on an Ubuntu Instance
3. Enabling Network Traffic to Ubuntu Images: Enabling Network Traffic to Ubuntu Images


Putting it altogether: How to deploy scalable and secure APEX on OCI

Oracle APEX is very popular, and it is one of the most common usecases that I see with my customers. Oracle Architecture Center offers a re...