Category Archives: Post-Mortem Analysis

Forensics Postmortem Analysis and how abusive bots pushed me to migrate part of my services from Google Cloud to another Cloud provider

I tested Google Cloud when they were in beta, years ago, and I helped their Team to improve the service. And I’ve been using Google Cloud for years.

Here I explain, how an incident, this August, almost made me leave Google Cloud for good. In fact, I migrated many services to other providers. I walk through the steps I took.

The best advice to start with, is, enable Cloud Billing alerts.

Everything started with the alert of Cloud Billing alerting me that half of the budget for the month had been exhausted around the 7th of the month. I checked the panel, and at the beginning I didn’t pay much attention to it, as Billing was indicating that the extra cost was in Compute in a specific region where I had provisioned new instances and more storage and it was just few euros.

Compute Engine Costs are mixed with the bandwidth consumption in the Dashboards which makes no sense:

So, I had to check different dashboards to realise it was a data transfer issue, and honestly, at the beginning I didn’t know where the problem was, but as I wrongly inferred, that the instance generating egress traffic was the one named “ubuntu26-04” I was investigating mainly that.

The first thing to learn that Google Cloud Billing shows the pure Compute and bandwidth usage together. I dislike it, for obvious reasons.

Next day budget for the month was almost gone.

Running the reports I saw “SKU Network Internet Data Transfer Out from Americas to Americas” which was confusing, and initially made me think if it could be the new backups system, sending the backups to my centralised server using the AI compatible agent with the public IP for the API and encryption, but I discarded this very quickly due to the volume of data and I modified the system to use internal IPs instead of the public ones.

The only clear thing at this point was that Internet data transfer (egress) was being sent, and Google makes you pay for that.

I had released several projects recently, like https://read.carlesmateo.com/en Text-to-speech to generate MP3 from text files, or https://sendmeafile.carlesmateo.com/en to send and receive huge files, with resume, checksum verification, AI compatible… So my first thought was that one of the projects or its dependencies could have been compromised, or fanatic AI/crawlers were downloading the sample MP3 files again and again.

I made a mistake when checking Cloud Billing. It indicated that the expenses came from us-central1, so checked the list of instances’ regions and I thought the problem came from a new instance named “ubuntu26-04”.

But I made a mistake, because there are several zones in every region, like us-central1-a, us-central1-b, us-central1-c… My new instance for the new projects was in the zone us-central1-f but I had more than that one instance in the region, only that in different zones. But when I checked the zones for the instances I did not realise that there was more than one instance in the zone us-central1 and I assumed the new instance was the only one in the zone and should be the one generating the egress traffic.

I only had ports 80 and 443 opened in Google’s firewall, so there could be requests to the web server, or that the server got hacked. I analysed the nginx logs for the server and found no huge traffic or big files being requested.

Nowadays bots and crawlers and AI’s generate a crazy amount of traffic, so I had to discard this.

Also, recently Amazon had a problem were it presented huge bills (erroneous) of more than 1 Million dollars to some customers. I didn’t know if Google was calculating the bandwidth costs well.

The Billing Dashboard showed that the increment in data sent was costing me around 11€ per day, so it allowed me to start moving services from that server, with relative calm. The forecast for all the August month displayed that the bill would cost me 300€, which was not nice.

In the meantime I hired a VPS with unlimited data and started to migrate the Docker services there as fast as I could with the idea of shutting down the compromise instance to cut the costs.

One of the things that really made me angry, while I was migrating to avoid those over costs, is that I checked again, the next day, the Cloud Billing Reports, the web loaded and informed me that it had changed with news and showed some floating popups pointing how it work and some improvements pushed by the Google Dev Team, and when I continued, the Cloud Billing Report was displaying an unexpected amount of money to pay in August 2026, more than 10 times what I expected, instead of the forecast I saw previously. So it looked like I had to pay that amount for August, and I started to look at what caused that increment of costs quickly, which fas frustrating. And after a while, when I was thinking about stopping the services, I found that Google had changed the page of the reports and the filters were reset to yearly, and the page was showing the already paid costs from 1st August 2025 to 31th August 2026.

Later I reported how frustrating was the experience to Google as “you should not do that” and “users dislike frequent changes”.

Companies are not aware that when they change things, they often annoy the customers a lot.

Years ago I migrated from my long term account with Amazon AWS to Google Cloud Compute Engine cause I was feed up of the problems in Amazon, how many programming errors they had, the changes they did, causing problems, like to the identifiers of the objects (I had older created objects, with shorter hash, and some API calls did not work any more), and how difficult they made simple things. User Interface was bad, and they also lost two of my instances and communicated really poorly (I also reported that to a friend in Amazon and I had some conversations with Amazon’s AWS PMs), and they were over expensive for developers, so I ended migrating to Google Cloud. And now Google is repeating some of the same mistakes.

They are not only expensive. They make things that could be easy, difficult, for the developers.

These big companies force you to stay continuously learning they changes. They offer great solutions for the Enterprise, but for developers and Startups the costs could be overkill.

Probably they are over-kill and over-expensive for developers and Startups.

With Google Cloud I have experienced some technical issues too, as my older instance froze several times and became irresponsive and I had to do a stop from the web console and I had to wait until the command timed out and Google realised there was a problem so it would consider the instance finally shutdown and I could start it again in other non faulty hardware.

If you are a company with a Load Balancer, with health checks, and several front servers, and 20 developers, you can afford to have Web servers KO, as the Load Balancer will handle it nicely, but if you have a single server with WordPress sites, if the instance stops it means that nobody can access those sites.

As an example of costs, for a Google e2-medium (2 vCPUs, 4 GB of Memory and 20 GB of disk) in us-central1 I pay around 30 € per month. And it happens that the forecast for this month in bandwidth was 300 €.

Also storage is around 1€/month per 10GB for standard. So 200GB of regular space is 20 €/month.

For SSD 1.7€/month per 10GB, even tough provides more IOPS, but not more speed. 200GB of space more is 34 € more per month.

I migrated to a provider were I pay 30 € per month for a 8 vCPU and 24 GB of memory and includes 200 GB of NVMe, and I have unlimited bandwidth. Extra storage is also much more cheap. And the static IP is included (for Google you pay around 1.1€ per month).

The key in here was the unlimited bandwidth.

I also could have rented a dedicated server with 64GB of RAM for around 64 €/month, but these comes with some down sides, like if your server experiences a hardware error, you’ll have to wait for the System Administrators from the Cloud Provider to physically replace the components. With a VPS or instance, I can just launch it in a healthy hardware immediately.

As I was migrating services and investigating the origin of the problems, I saw that 500 GB of data were transferred from us-central1 since the 1st of August.

The Billing Report didn’t allow me to filter for IP. It only helped to know the region, us-central1, which helped me to identify the possible services that could be having the problems.

To be honest, I should have located the origin of the problem quickly, but I did so well blocking bots and uneducated crawlers in the past years, that I had no problems for years, and GCE has mad many changes over time, so I did not remember that I could monitor real time network usage of my servers. Most of the servers I troubleshoot I have only SSH access, so I use network tools.

But as I told previously, I mistakenly thought that just one instance was running in us-central1, and that was not the case. They were in different zones of us-central1.

The live traffic inspection in the server with tools like ss, netstat, nethogs, iftop, iotop and ps revealed normal traffic and load.

A second revision from the nginx logs didn’t reveal any problem over time. And the only opened ports were 80 and 443. With the services I considered potentially more vulnerable (as they used python dependencies) migrated to the new VPS and stopped in the “ubuntu26-04” Google server, the usage shown and the forecast for bandwidth was not reduced and it was the same.

One thing I learned, and I suspected it from my tests with https://sendmeafile.carlesmateo.com/en is that the maximum speed that my instances were able to transfer from Google to Internet (egress) was around 50 Mbps per connection.

I didn’t see this in Google Cloud’s documentation, but in my tests transferring data from Google servers using the public interface and from different locations to this google server I consistently achieved always that top speed of around 50 Mbps per connection (around 6.25 MB/s), but when I communicated Google servers using their private interfaces, the speed for a single connection was much more higher. So I assumed that the bandwidth was capped. After seeing the data transferred in the Billing Cloud Report, it looks clear to me that an important part of the server’s bandwidth max was being used. Later, Apache logs, confirmed me many requests at the same exact second.

Google anomalies tool detected nothing despite the huge increment in data transfer and cost variation (8,960% respect July!).

Before taking the decision to migrating to another server one has to think, not only in costs, but also on how trustworthy are the storage systems, the resiliency of the systems, how many times they are down per year… there are surprises that you discover as you go. With the experience.

But honestly, I felt it was not worth the price to continue with Google Cloud and it was conditioning the kind of projects I could start. My architecture is resilient, based in containers and I can deploy it fast. I also have backups.

Google storage is not that super trustable in my opinion (neither Amazon’s). Some times my VMs froze and Google monitor dashboard stops drawing the lines of metrics (like CPU usage), but Google doesn’t catch that the VM is frozen. My guess is that the problem is the storage or compute servers dying.

Having to pay per GB transferred in a time where bots and AI index your sites continuously, and abuse the servers by launching many requests concurrently, generating load, and where bots consume much more data than humans, force you to take architecture decisions and to implement strategies.

Possible solutions like using CDN’s like Cloudflare to be protected against DoS, to cache statics and to cut bandwidth and CPU abuse (requests to origin), using other cache systems or migrating to other Cloud providers that are more developer and Start up friendly, and that have unlimited data transfer, etcetera.

When I was presented with the new Google Billing Reports page and the filter was reset to one year, and I saw the price that should be showing multiplied by 10, I contacted Google Cloud by X. They kindly replied and invited me to contact them privately. I did and we opened a ticket. Around three hours later I got an email from a Google staff, telling me that they will investigate the problem, and giving me an expected response deadline of 5 days.

It was nice, but by the time I received the email I already had found and sorted the problem.

I troubleshoot the instance “ubuntu26-04” but the CPU usage was low, and no permanent connections (a sign that could have indicated that was hacked):

That Looked normal

sudo ss -tunp state established

Then I did something that I should have done before, I checked the metrics from all the different instances in Google Cloud Compute, starting from the one I initially guessed was the one suffering the problems. That machine was clear. No high CPU usage. No high bandwidth usage. Nothing.

And then I saw it.

Through VM Instances I narrowed to the one in us-central1 that was consuming the Internet bandwidth, “instance-1”:

I was checking the wrong server. The consuming excess of egress was from instance-1 that hosted several WordPress sites and it was using Apache2, not nginx.

Some commands I ran:

# Check for the data sent by Apache. This found a lot of traffic. Note: the access.log of all the sites were being analysed
sudo awk '{s+=$10} END {print s/1024/1024/1024 " GB"}' /var/log/apache2/*access.log
# returned 36.4578 GB that's for a day, so something was clearly wrong

# Analyze requests: See what is more heavy. So basically bots trying to hack into the server and some heavy images. Note root home for at least a server is huge (50MB)
sudo awk '{a[$7]+=$10} END {for(u in a) print a[u]/1024/1024" MB\t"u}' /var/log/apache2/*access.log | sort -rn | head -20
54.3218 MB      /
21.6173 MB      /wp-login.php
17.1676 MB      /wp-content/uploads/2022/10/20221001_200930-cut.png
16.5822 MB      /wp-content/uploads/2022/10/20221001_201033.png
10.7245 MB      /wp-content/uploads/2022/02/python_loop.gif
6.61174 MB      /xmlrpc.php
6.60595 MB      /page/2/?shopdetail%2FI37266592159
6.60595 MB      /page/2/?shopdetail%2FI37256267952
6.60595 MB      /page/2/?shopdetail%2FI37256137878
6.60595 MB      /page/2/?shopdetail%2FI37256033568
6.60595 MB      /page/2/?shopdetail%2FI37224959793
6.60595 MB      /page/2/?shopdetail%2FI37214305672
6.60595 MB      /page/2/?shopdetail%2FI37203620584
6.60595 MB      /page/2/?shopdetail%2FE37266925367
6.60595 MB      /page/2/?shopdetail%2FE37242498260
6.60595 MB      /page/2/?shopdetail%2FE37207321261
6.60595 MB      /page/2/?shopdetail%2FA37259909717
6.60595 MB      /page/2/?shopdetail%2FA37259600135
6.60595 MB      /page/2/?shopdetail%2FA37256306882
6.60595 MB      /page/2/?shopdetail%2FA37256197883

# Analyze requests: See what IP's are requesting more traffic in total. Gotcha
sudo awk '{a[$1]+=$10} END {for(i in a) print a[i]/1024/1024" MB\t"i}' /var/log/apache2/*access.log | sort -rn | head -20
34549.2 MB      74.7.241.60
1491.36 MB      116.202.235.23
287.651 MB      3.41.188.33
232.362 MB      74.7.227.160
199.987 MB      18.97.9.169
183.748 MB      66.249.74.42
84.1681 MB      114.119.140.64
78.3209 MB      114.119.149.169
70.5166 MB      114.119.156.46
67.8224 MB      114.119.150.29
66.7544 MB      114.119.145.237
58.7278 MB      114.119.147.113
57.1822 MB      114.119.162.62
56.0061 MB      66.249.74.43
55.3612 MB      114.119.141.79
21.9245 MB      216.73.217.35
21.2516 MB      104.243.44.93
17.7313 MB      79.116.217.181
17.1676 MB      189.81.71.184
16.5822 MB      205.188.63.181

So we see that this IP 74.7.241.60 is requesting 34.5 GB in a day.

So, according to this IP Info service, the IP is from Microsoft. So apparently, crawlers/IA from Microsoft or running in Microsoft infrastructure, were costing me money because they are abusing, scanning my sites, with many connections concurrently, like horrible soulless machines:

I double checked the IP with another provider:

A command to see a sample of what kind of requests I have from this IP, brings more light:

grep "74.7.241.60" /var/log/apache2/*access.log | head -20
/var/log/apache2/www_mysqlproxycache_com-access.log:74.7.241.60 - - [13/Aug/2026:00:00:22 +0000] "GET /page/2/?shopdetail%2FI37207088742 HTTP/1.1" 200 3463409 "http://www.mysqlproxycache.com/page/2/?shopdetail/I37207088742" "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; GPTBot/1.4; +https://openai.com/gptbot)"
/var/log/apache2/www_mysqlproxycache_com-access.log:74.7.241.60 - - [13/Aug/2026:00:00:34 +0000] "GET /page/3/?shopdetail/I37207088742 HTTP/1.1" 301 264 "http://www.mysqlproxycache.com/page/2/?shopdetail/I37207088742" "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; GPTBot/1.4; +https://openai.com/gptbot)"
/var/log/apache2/www_mysqlproxycache_com-access.log:74.7.241.60 - - [13/Aug/2026:00:00:35 +0000] "GET /page/3/?shopdetail%2FI37207088742 HTTP/1.1" 200 1900262 "http://www.mysqlproxycache.com/page/3/?shopdetail/I37207088742" "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; GPTBot/1.4; +https://openai.com/gptbot)"
/var/log/apache2/www_mysqlproxycache_com-access.log:74.7.241.60 - - [13/Aug/2026:00:00:37 +0000] "GET /page/7/?shopdetail/E37249347969 HTTP/1.1" 301 264 "http://www.mysqlproxycache.com/page/6/?shopdetail/E37249347969" "Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; GPTBot/1.4; +https://openai.com/gptbot)"

According to the User-Agent is GPTBot. So apparently a bot running in Microsoft self-identifying as OpenAI were killing causing high usage to my server and to my budget by requesting many pages at the same time, for an old shopdetail page that does not exist in that server (maybe there was a shop years before I got that IP), and spends a lot of bandwidth responding with a heavy page of 6 MB.

In the logs I saw many request at the same precise second.

Asking to Claude it tells me that the range of IPs of the Microsoft attacker IP are not in the publicly published IPs of OpenAI for GPTBot, so apparently somebody using Microsoft IPs is pretending to be a GPT Bot, and it isn’t:

The domain is an old WordPress that has been hacked. Yes, this happens more often that we would like, with WordPress.

I checked the https and it returns a SSL potential warning in the SSL.

I tried with http http://www.mysqlproxycache.com/page/7/?shopdetail%2FI37224744693 (one of the requests in the logs) and I saw that the WordPress site was hacked.

So first things first: I stopped Apache.

sudo service apache2 stop

So, to summarise, there was a bot from a Microsoft IP, that was requesting many times per second pages to a WordPress site that was compromised and it was returning a large article with many photos, so heavy in terms of data usage.

With Apache2 server stopped, I moved the config symbolic link from /etc/apache2/sites-enabled/ to another folder /etc/apache2/sites-disabled/ that I created.

So basically, when I started Apache, the compromised site will not be served. It will be served by the default site.

I restarted the Apache server and checked that the URLs for that domain were catch by default site and no processing resources of bandwidth was used.

That was to avoid a main site marked as default in the Apache config to attend those requests and waste data.

And then I added to the firewall, to Google Cloud’s Firewall and to the servers in all my instances, to block all the IPs from Microsoft from that range.

I saw that the BGP Prefix is 74.7.0.0/16 so I block all the 65,536 IP Addresses:

ufw insert 1 deny from 74.7.0.0/16 to any

As you can see I inserted this rule on the first position. Don’t just add it with ufw deny from 74.7.0.0/16 to any or otherwise it will go to the last position, after the allow rules. And then it would not be enforced.

Check your rules with:

ufw status numbered

You can block the range in your provider’s Firewall (VPC Firewall in Google Cloud).

If you didn’t have the Ubuntu Firewall active, and want to activate it now, make sure that you have the rules for the SSH connection:

echo "Enabling SSH"
sudo ufw allow 22/tcp

echo "Enabling Web 443 and 80"
#sudo ufw allow "Apache Full"
sudo ufw allow 80/tcp
sudo ufw allow 443/tcp

echo "Enabling Firewall"
sudo ufw enable

echo "Showing rules"
sudo ufw status numbered

I recommend you to add the rules to a script in git (Infrastructure as Code), so you can replay it another day, and to any server you have.

By default, ufw sets deny by default for rules not specified. You can check it with:

sudo ufw status verbose | head

Bots, IA, crawlers abuse our servers nowadays.

Also the automated bots attempting to hack into the servers.

If you just check your logs, you’ll see many attempts to exploit vulnerabilities in your servers, every day.

I checked for the second IP consuming more bandwidth in the list:

116.202.235.23

It’s from hetzner, a Cloud provider. https://bgp.he.net/net/74.7.0.0/16

Another 65,536 addresses that get blocked:

# Hetzner
ufw insert 1 deny from 116.202.0.0/16  to any

The third IP will be blocked too, even if I didn’t find any information about who owns the block of IPs.

ufw insert 1 deny from 3.41.188.33 to any

The fourth was part of the same block I blocked first, so it’s already blocked.

The fifth IP, I ask to Google Gemini:

The block is 18.97.9.0/24 but I’ll block all the BGP Prefix, all the 16K addresses.

ufw insert 1 deny from 18.97.0.0/18 to any

See what is the traffic from the servers since I blocked the Microsoft range in the Ubuntu Firewall:

As you can see, after I blocked the first abusive IP allegedly from Microsoft with seemingly OpenAI User-Agent falsified, all the egress traffic went down.

I like how Google reacted when I contacted them. And I should have checked the Network traffic bandwidth first. So I decided to keep the WordPress instance in Google Cloud.

But the new instance “ubuntu26-04” I’m migrating it to the new server. What they offer for the price is much better, and the price of the storage really it’s important for some of my projects. The unlimited bandwidth usage in the new provider it’s also key.

The Billing report from Google Cloud doesn’t show the data up to the moment. But the 16th of August I was able to see that from the 12th and absolutely on 13th and 14th the egress Internet cost have almost disappeared.

Not all is perfect. I had to learn the web interface from that provider (I used it in the past, but companies change the user interface very often nowadays, causing headaches to users), the place were my VPS was listed was hidden and difficult to find, and by default the VPS had no firewall enabled. Something that I believe that is basic to have enabled by default.

But I think it will be great for my projects and for my budget, to have all those new resources available (8 vCPU and 24 GB of RAM and 200 GB of NVMe disk for the same cost provided in Google Cloud of 2 vCPU and 4 GB of RAM and 20 GB of standard disk).

In fact, the speed I have detected transferring from/to sendmeafile with the new provider shows a great improvement.

Take one thing in mind: They AI is hungry for contents, and the mega-rich AI providers are making you pay the bill, by scanning your sites non-stop.

The new provider says it provides 3 Gbps of bandwidth for the instance, with no data usage limits (not having to pay for GB transferred). If I had not reacted fast, the forecast for this month with Google for just a single instance was 300€. It is a risk for developers and small companies, to have the possibility to face a huge bill because more frequent every day bots/crawlers/AIs and processes attempting to breach in your servers, cause your server to use a lot of egress data. So it is worth considering using a Cloud Provider with unlimited data plan. Or at least, activate the Billing Alerts to be notified if your costs exceed a certain threshold.

If you have problems with abusive bots degrading your Drupal or WordPress performance, I created a commercial WordPress plugin and a commercial Drupal module that automatically and easily record and block offending IP Addresses.

Example of a Claude Fable 5 error letting my payment process down

As I’m seeing many people that embraces the hype of blindly trusting IA, and not reviewing the code, I want to share with you a case, where Claude Fable 5 introduced an error in Production, in the payment process of a side project, in July 2026.

Reviewing routinely the payment transactions, I noticed a 100€ purchase from which I didn’t know.

I located the user that paid and I saw that in the database that payment was not registered.

I checked the PHP logs in that time range and I saw a 500 error, just after the user paid in the payment provider and came back to the site, just in the moment that the database is updated.

The user tried 3 times to refresh the payment page.

The error was, that Claude had introduced objects of the type DateTimeImmutable, so a date/time that cannot be changed.

But Doctrine was mapping a a mutable DateTime, and that was incompatible, and crashed.

It was a programming error introduced by Claude.

So, after a successful payment, when the user came back to the site, it would get an error, always, and the table reflecting that the plan was paid would never get updated.
I didn’t review most of the code of the new features. When I created the project I tested against the test sandbox and everything worked, and I moved to Production. But at some point, in the development of the next features, the payment process was broken.

So, Claude broke the payment, and neither me or it detected it.

After every new code change I had tried the new functionalities, checked the database changes, but I didn’t validate again the payment flow. Basically, after I moved to Production my provider disabled the Test payment sandbox, so, in order to test the full flow I had to pay 100€ for real and as I tried at the beginning, I didn’t try it again.

So, at some point, Claude introduced a bug, that I didn’t detect, that Claude didn’t detect, and that was impacting directly in the payment process, and I had a user that paid and that had no service and he never contacted me.

I reached the user and explained that I had a technical problem, and I offered a refund but the user declined, he wanted the service. What I did was to upgrade their plan for free to have 100 times more capacity, as thanks for their fidelity.

I implemented the mechanisms so an error like this cannot happen again, but as I see many decision makers trusting blindly the AI, I wanted to share this real experience, not detected, with impact in the payment process.

This is a good lesson to learn with a personal side project that it’s starting, but would be an ugly one to discover with an already running system from a system, already generating thousands of euros per hour.

That’s the explanation that Claude Fable 5 provided to me.

If you use AI for generating code you know that it makes mistakes. Claude is very good, but some times it makes mistakes. It may be a random character introduced in a Varnish config file, or a missing + sign, pointing to an inexistent line number on a file, or introducing incorrect expression in a nginx host config section or a Symfony route. Most of the times the modern models catch the errors before, but not always.

We spent decades trusting deterministic programs, which given the same input and environment, produce the same output.

But you should bear in mind that the generative AI is not deterministic, which means that given the same input it produces different outputs.

Having different agents with different roles is useful, and an improvement, although it’s expensive in tokens and may miss errors.

Asking for a Code Review to the AI is a good thing too. But some times, the AI does not follow the rules, and ignores your instructions and not everything is cached.

That’s why I recommend to use deterministic processes, to add more certainty: Automations, programs and scripts that validate, Unit Testing, Functional testing, etc…

Migrating my 11 years Amazon AWS account services (Postmortem Analysis)

I started to explain that I was migrating some services from Amazon and that some of my sites were under Maintenance and that I would provide more information.

Here is the complete history of why I migrated all the services from my 11 years old Amazon account to other CSP.

Some lessons can be learned from my adventure.

I migrated my last services from Amazon to GCP

Amazon sent me an email on October 6th, this year 2021, telling me that they will disable EC2-Classic by August 2022. I thought I would not be able to keep my Static Ip’s as in the past VPC Ip’s and EC2-Classic Ip’s were not transferable, so considering that I would loss my Static Ip’s anyway I started to migrate to some to other providers like Digital Ocean.

Is not cool losing Static Ip (Elastic Ip in AWS) Addresses as this is bad for SEO, so given that I though I would lose my Static Ips that have been with me for years, I started to migrate certain services to providers much more economic.

Amazon is terrible communicating, and I talked with some product managers in the past about that, when they lost one of my Volumes, and the email was so cold and terrible that actually that hurt more than Amazon losing my Data. I believed that it was a poorly made Scam and when I realized it was true I reached one of my friends, that is manager there, as I know they care for doing things right, and he organized a meeting with two PM so I can pass my feedback.

The Cloud providers are changing things very fast, and nobody is able to be up to date with the changes, unless their work position allows plenty of time to get updated. Even if pages of documentation are provided, you have to react to an event that they externally generated forcing you to action. Action to read all the documentation about EC2-Classic migrations, action to prepare to have migrated by August 2022.

So August 2022… I was counting that I had plenty of time but I’m writing a new book about using the Amazon SDK for Python, boto3, and I was doing some API calls and they started to fail in a very unusual way, Exceptions with timeout, but only for the only region where I had EC2-Classic. 

urllib3.exceptions.NewConnectionError: <botocore.awsrequest.AWSHTTPSConnection object at 0x7f0347d545e0>: Failed to establish a new connection: [Errno -2] Name or service not known

My config was:

        o_config = Config(
            region_name="us-east-1a",
            signature_version="v4",
            retries={
                'max_attempts': 10,
                'mode': 'standard'
            }
        )

But if I switched to another region name, it would work:

            region_name='us-west-2',

I made a mistake in here, the region name is “us-east-1” and not “us-east-1a“. “us-east-1a” is the availability zone. So the SDK was giving a timeout because in order to connect to the endpoint it uses the region name as part of the hostname. So it doesn’t find that endpoint because it doesn’t exist.

I never understood why a company like Amazon is unable to provide the SDK with a sample project or projects 100% working, with the source code so people has a base that works to build up.

Every API that I have created, I have provided it with documentation but also with example for several languages for how to use it.

In 2013 I was CTO of an online travel agency, and we had meta-searchers consuming our API and we were having several hundreds of thousands requests per second. Everything was perfectly documented, examples were provided for several languages, the document and the SDK had version numbers…

Everybody forgets about Developers and companies throw terrible and cold products to the poor Developers, so difficult to use. How many Developers would like to say: Listen Mr. President of the big Cloud Company XXXX, I only want to spawn a VM that works, and fast, with easy wizards. I don’t want to learn 50 hours before being able to use your overpriced platform, by doing 20 things before your Ip’s are reflexes of your infrastructure and based in Microservices. Modern JavaScript frameworks can create nice gently wizards even if you have supercold APIs.

Honestly, I didn’t realize my typo in the region and I connected to the Amazon Console to investigate and I saw this.

Honestly, when I read it I understood that they were going to end my EC2 Networking the 30th of October. It was 29th. I misunderstood.

It was my fault not reading it well to the end, I got shocked by the first part telling about shutdown and I didn’t fully understood as they were going to shutdown EC2-Classic for the zones I didn’t had anything running only.

From the long errors (3 exceptions chained) I didn’t realize that the endpoint is built with the region name. (And I was passing the availability zone)

botocore.exceptions.EndpointConnectionError: Could not connect to the endpoint URL: "https://ec2.us-east-1a.amazonaws.com/"

Here is when I say that a good SDM would had thought and cared for the Developers more, and would had made the SDK to check if that region exists. How difficult is to create a SDK a bit more clever that detects a invalid region id?. It is not difficult.

It is true that it was late in the evening and I was tired of all the day, and two days of the week between work and zoom university classes I work 15 hours and 13 hours respectively, not counting the assignments, so by the end of the week I am very tired. But that’s why it is very important to follow methodology and to read well. I think Amazon has 50% of the fault by the way they do things: how the created the SDK, how they communicate, and by the errors that the console returned me when I tried to create a VPC instance of an EC2-Classic AMI (they seem related to the fact I had old VPC Network objects with shorter hash than the current they use) and the other 50% was my fault for not identifying the source of the error, and not reading the message in their website well.

But the fact that there were having those errors in the API’s and timeouts made me believe they were going to cut the EC2-Classic Networking the next day.

All the mistakes fall together in a perfect storm.

I checked for documentation and I saw it was possible to migrate my Static Ip’s to VPC Static Ip’s.

It was Friday evening, and I cancelled my plans, in order to migrate the Blog to VPC in an attempt to keep running it with Amazon.

As Cloud Architect, I like to have running instances in several CSP as it allows me to stay up to date with the changes they do.

I checked the documentation for the migration. Disassociating the Static Ip (Elastic Ip in AWS jargon) was easy. Turning into VPC as well.

As I progressed, what had to be easy turned into a nightmare, as I was getting many errors from the Amazon API, without any information, and my Instances were not created.

I figured out that their API could have problems with old VPC objects I created time ago, so I had to create new objects for several things.

I managed to spawn my instances but they were being launch and terminated instantly without information. Frustrating.

When launching a new instance from the AMI (a Snapshot of the blog), I was giving shown options to add more volumes without any sense. My Instance was using 16GB from a 20GB total Space, and I was shown different volume configs, depending on the instance, in some case an additional 20GB volume, in other small SSD, ephemeral and 10 GB for the AMI (which requires at least 16GB).

After some fight I manage to make it work after deleting the volumes that made no sense, and keeping only one of 20GB, the same size of my AMI.

But then my nightmare started to make the VPC Instance to have Internet access and to be seen from outside. I had to create a new Internet Gateway, NAT, Network, etc…

As mentioned the old objects I was trying to reusing were making the process to fail.

I was running out of time, and I thought in few time they were going to shutdown EC2-Classic network (as I did not read correctly), so I decided to download everything and to migrate to another provider. For doing that first I blocked all the traffic, except for my Ip.

I worked in parallel, creating the new config in Google Cloud, just in case I had forgot something. I had created a document for the migration and it was accurate.

I managed to do everything fast enough. The slower part was to download all the Data, as I hold entire VM’s for projects like Cassandra Universal Driver.

Then I powered off my Amazon Instance for the Blog forever.

In GCP I blocked all the traffic in the firewall, except for my Ip, so I could work calmly.

When everything was ready, I had to redirect the DNS to the new static Ip from Google.

The DNS provider I used had implemented some changes in their API so I was getting errors replacing my old entry ‘.’ (their JSON calls returned Internal Server Error). Finally I figured it out how to workaround it and I was able to confirm that the first service was up and running.

I did some tests to make sure there were not unexpected permission problems, entries in the logs, etc…

Only then I opened the Google Firewall. I have a second firewall in each instance where I block or open at Ip tables level what I want. Basically abusive bot’s IPs trying to find exploits or brute force by dictionary passwords.

I checked with my phone, without Wifi that the Firewall was all good. (It is always a good idea to use another external Ip, different from the management one, to check)

I added a post explaining that I was migrating some of my Services and were under maintenance.

I mentioned in the blog that some of my services were being migrated from Amazon to Digital Ocean.

For some reasons, in the Backup of the Database one user was lost, so I created it in the MySQL with the typical commands:

CREATE USER 'username'@'localhost' IDENTIFIED BY 'password';
FLUSH PRIVILEGES;
GRANT ALL PRIVILEGES ON mydatabase.* TO 'username'@'localhost';

Swap, swappiness, Servers not responding

I have read a lot of wrong recommendations about the use of Swap and Swappiness so I want to bring some light about it.

The first to say is that every project is different, so it is not possible to make a general rule. However in most of the cases we want systems to operate as fast and efficiently as possible.

So this suggestions try to covert 99% of the cases.

By default Linux will try to be as efficient as possible. So for example, it will use Free Memory to keep IO efficient by keeping in Memory cache and buffers.

That means that if you are using files often, Linux will keep that information cached in RAM.

The swappiness Kernel setting defines what tradeoff will take Linux between keeping buffers with Free Memory and using the available Swap Memory.

# sysctl vm.swappiness
vm.swappiness = 60

The default value is 60 and more or less means that when RAM memory gets to 60%, swap will start to be used.

And so we can find Servers with 256GB of RAM, that when they start to use more than 153 GB of RAM, they start to swap.

Let’s analyze the output of free -h:

carles@vbi78g:~/Desktop/Software/checkswap$ free -h
              total        used        free      shared  buff/cache   available
Mem:          2.9Gi       1.6Gi       148Mi        77Mi       1.2Gi       1.1Gi
Swap:         2.0Gi        27Mi       2.0Gi

So from this VM that has 2.9GB of RAM Memory, 1.6GB are used by applications.

The are 148MB that can immediately used by Applications, and there are 1.2GB in buffers/cache. Does that means that we can only use 148MB (plus swap)?. No, that mean that Linux tried to optimize io speed by keeping 1.2GB of RAM memory in buffers. But this is the best effort of Linux to have performance, for real applications will be also able to use 1.1GB that corresponds to the available field.

About swap, from 2GB, only 27MB have been used.

As vm.swappiness is set to 60, more RAM will be swapped out to swap, even if we have lots available.

As I said every case is different. If we are talking about a Desktop that has NVMe drives, the impact will be low. But if we are talking about a Server that is a hypervisor running VMs and has high usage on CPU and has the swap partition or the swap in a file, that could lead to huge problems. If there is a physical Server with a single spinning drive (or logical unit through RAID), and one partition is for Swap, and the other for mountpoints, and a process is heavily reading/writing to a partition mounted (an elastic search, or a telegraf, prometheus…), and the System tries to swap, then they will be competing for the magnetic head of disk, slowing down everything.

If you take a look on how the process of swapping memory pages from the memory to disk, you will understand that applications may need certain pages before being able to run, so in many cases we get to lock situations, that force everything to wait.

In my career I found Servers that temporarily stopped responding to ping. After a while ping came back, I was able to ssh and uptime showed that the Server did not reboot.

I troubleshooted that, and I saw a combination of high CPU usage spikes and Swap usage.

Using iostat and iotop I monitored what was speed of transference of only 1 MB/second!!.

I even did swapoff and it took one hour to free 4 GB swap partition!.

I also saw swap partition being in a spinning disk, and in another partition of the same spinning drive, having a swapfile. Magnetic spinning drives can only access one are of the drive at the same time, so that situation, using swap is very bad.

And I have seen situations were the swap or swapfile was mounted in a block device shared via network with the Server (like iSCSI or NFS), causing terrible performance when swapping.

So you have to adapt the strategy according to the project.

My preferred strategy for Compute Nodes and NoSQL Databases is to not use swap at all. In other cases, like MySQL Databases I may set swappiness to preferably to 1 or to 10.

I quote here the recommendations from couchbase docs:

The Linux kernel’s swappiness setting defines how aggressively the kernel will swap memory pages versus dropping pages from the page cache. A higher value increases swap aggressiveness, while a lower value tells the kernel to swap as little as possible to disk and favor RAM. The swappiness range is from 0 to 100, and most Linux distributions have swappiness set to 60 by default.

Couchbase Server is optimized with its managed cache to use RAM, and is capable of managing what should be in RAM and what shouldn’t be. Allowing the OS to have too much control over what memory pages are in RAM is likely to lower Couchbase Server’s performance. Therefore, it’s recommended that swappiness be set to the levels listed below.

https://docs.couchbase.com/server/current/install/install-swap-space.html

Another theme, is when you log to a Server and you see all the Swap memory in use.

Linux may have moved the pages that were less used, and that may be Ok for some cases, for example a Cron Service that waits and runs every 24 hours. It is safe to swap that (as long as the swap IO is decent).

When Kernel Swaps it may generate locks.

But if we log to a Server and all the Swap is in use, how can we know that the Swap has been quiet there?.

Well, you can use iostat or iotop or you can:

cat /proc/vmstat

This file contains a lot of values related to Memory, we will focus on:

pswpin 508992338
pswpout 280871088

In https://superuser.com/questions/785447/what-is-the-exact-difference-between-the-parameters-pgpgin-pswpin-and-pswpou you can find very interesting description of those values. I paste here an excerpt:

Paging refers to writing portions, termed pages, of a process’ memory to disk.
Swapping, strictly speaking, refers to writing the entire process, not just part, to disk.
In Linux, true swapping is exceedingly rare, but the terms paging and swapping often are used interchangeably.

page-out: The system’s free memory is less than a threshold “lotsfree” and unnused / least used pages are moved to the swap area.
page-in: One process which is running requested for a page that is not in the current memory (page-fault), it’s pages are being brought back to memory.
swap-out: System is thrashing and has deactivated a process and it’s memory pages are moved into the swap area.
swap-in: A deactivated process is back to work and it’s pages are being brought into the memory.

Values from /proc/vmstat:

pgpgin, pgpgout – number of pages that are read from disk and written to memory, you usually don’t need to care that much about these numbers

pswpin, pswpout – you may want to track these numbers per time (via some monitoring like prometheus), if there are spikes it means system is heavily swapping and you have a problem.

In this actual example that means that since the start of the Server there has been 508992338 Page Swap In (with 4K memory pages this is 1,941 GB, so almost 2 TB transferred) and for Page Swat Out (with 4K memory pages this is 1,071 GB, so 1 TB of transferred). I’m talking about a Server that had a 4GB swap partition in a spinning disk and a 12 GB swapfile in another ext4 partition of the same spinning disk.

The 16 GB of swap were in use and iotop showed only two sources of IO, one being 2 VMs writing, another was a journaling process writing to the mountpoint where the swapfile was. That was an spinning drive (underlying hardware was raid, for simplicity I refer to one single drive. I checked that both spinning drives were healthy and fast). I saw small variations in the size of the Swap, so I decided to monitor the changes in pswpin and pswpout in /proc/vmstat to see how much was transferred from/to swap.

I saw then how many pages were being transferred!.

I wrote a small Python program to track those changes:

https://gitlab.com/carles.mateo/checkswap

This little program works in Python 2 and Python 3, and will show the evolution of pswpin and pswpout in /proc/vmstat and will offer the average for last 5 minutes and keep the max value detected as well.

As those values show the page swaps since the start of the Server, my little program, makes the adjustments to show the Page Swaps per second.

A cheap way to reproduce collapse by using swap is using VirtualBox: install an Ubuntu 20.04 LTS in there, with 2 GB of less of memory, and one single core. Ping that VM from elsewhere.

Then you may run a little program like this in order to force it to swap:

#!/usr/bin/env python3
a_items = []
i_total = 0
# Add zeros if your VM has more memory
for i in range(0, 10000000):
    i_total = i_total + i
    a_items.append(i_total)

And checkswap will show you the spikes:

Many voices are discordant. Some say swappiness default value of 60 is good, as Linux will use the RAM memory to optimize the IO. In my experience, I’ve seen Hypervisors Servers running Virtual Machines that fit on the available physical RAM and were doing pure CPU calculations, no IO, and the Hypervisor was swapping just because it had swappiness to 60. Also having swap on spinning drives, mixing swap partition and swapfile, and that slowing down everything. In a case like that it would be much better not using Swap at all.

In most cases the price of Swapping to disk is much more higher than the advantage than a buffer for IO brings. And in the case of a swapfile, well, it’s also a file, so my suspect is that the swapfile is also buffered. Nothing I recommend, honestly.

My program https://gitlab.com/carles.mateo/checkswap may help you to demonstrate how much damage the swapping is doing in terms of IO. Combine it with iostat and iotop --only to see how much bandwidth is wasted writing and reading from/to swap.

You may run checkswap from a screen session and launch it with tee so results are logged. For example:

python3 checkswap.py | tee 2021-05-27-2107-checkswap.log

If you want to automatically add the datetime you can use:

python3 checkswap.py | tee `date +%Y-%m-%d-%H%M`-checkswap.log

Press CTRL + a and then d, in order to leave the screen session and return to regular Bash.

Type screen -r to resume your session if this was the only screen session running in background.

An interesting reflection from help Ubuntu:

The “diminishing returns” means that if you need more swap space than twice your RAM size, you’d better add more RAM as Hard Disk Drive (HDD) access is about 10³ slower then RAM access, so something that would take 1 second, suddenly takes more then 15 minutes! And still more then a minute on a fast Solid State Drive (SSD)…

https://help.ubuntu.com/community/SwapFaq

Do you have a swap history that you want to share?.

Adding a swapfile on the fly as a temporary solution for a Server with few memory

Here is an easy trick that you can use for adding swap temporarily to a Server, VMs or Workstations, if you are in an emergency.

In this case I had a cluster composed from two instances running out of memory.

I got an alert for one of the Servers, reporting that only had 7% of free memory.

Immediately I checked it, but checked also any other forming part of the cluster.

Another one appeared, had just only a bit more memory than the other, but was considered in Critical condition too.

The owner of the Service was contacted and asked if we can hold it until US Business hours. Those Servers were going to be replaced next day in US Business hours, and when possible it would be nice not to wake up the Team. It was day in Europe, but night in US.

I checked the status of the Server with those commands:

# df -h

There are 13GB of free space in /. More than enough to be safe as this service doesn’t use much.

# free -h
              total        used        free      shared  buff/cache   available
Mem:           5.7G        4.8G        139M        298M        738M        320M
Swap:            0B          0B          0B

I checked the memory, ok, there are only 139MB free in this node, but 738MB are buff/cache. Buff/Cache is memory used by Linux to optimize I/O as long as it is not needed by application. These 738 MB in buff/cache (or most of it) will be used if needed by the System. The field available corresponds to the memory that is available for starting new applications (not counting the swap if there was any), and basically is the free memory plus a fragment of the buff/cache. I’m sure we could use more than 320MB and there is a lot if buff/cache, but to play safe we play by the book.

With that in mind it seemed that it would hold perfectly to Business hours.

I checked top. It is interesting to mention the meaning of the Column RES, which is resident memory, in other words, the real amount of memory that the process is using.

I had a Java process using 4.57GB of RAM, but a look at how much Heap Memory was reserved and actually being used showed a Heap of 4GB (Memory reserved) and 1.5GB actually being used for real, from the Heap, only.

It was unlikely that elastic search would use all those 4GB, and seemed really unlikely that the instance will suffer from memory starvation with 2.5GB of 4GB of the Heap free, ~1GB of RAM in buffers/cache plus free, so looked good.

To be 100% sure I created a temporary swap space in a file on the SSD.

(# means that I’m executing this as root, if you type literally with # in front, this will be a comment)

# fallocate -l 1G /swapfile-temp

# dd if=/dev/zero of=/swapfile-temp bs=1024 count=1048576 status=progress
1034236928 bytes (1.0 GB) copied, 4.020716 s, 257 MB/s
1048576+0 records in
1048576+0 records out
1073741824 bytes (1.1 GB) copied, 4.26152 s, 252 MB/s

If you ask me why I had to dd, I will tell you that I needed to. I checked with command blkid and filesystem was xfs. I believe that was the reason.

The speed writing to the file is fair enough for a swap.

# chmod 600 /swapfile-temp

# mkswap /swapfile-temp
Setting up swapspace version 1, size = 1048572 KiB
no label, UUID=5fb12c0c-8079-41dc-aa20-21477808619a

# swapon /swapfile-temp

I check that memory is good:

# free -h
              total        used        free      shared  buff/cache   available
Mem:           5.7G        4.8G        117M        298M        770M        329M
Swap:          1.0G          0B        1.0G

And finally I check that the Kernel parameter swappiness is not too aggressive:

# sysctl vm.swappiness
vm.swappiness = 30

Cool. 30 is a fair enough value.

2022-01-05 Update for my students that need to add additional 16GB of swap to their SSD drive:

sudo fallocate -l 16G /swapfile-temp
sudo dd if=/dev/zero of=/swapfile-temp bs=1024 count=16777216 status=progress
sudo chmod 600 /swapfile-temp
sudo mkswap /swapfile-temp
sudo swapon /swapfile-temp

Post-Mortem: The mystery of the duplicated Transactions into an e-Commerce

Me, with 4 more Senior BackEnd Engineers wrote the new e-Commerce for a multinational.

The old legacy Software evolved into a different code for every country, making it impossible to be maintained.

The new Software we created used inheritance to use the same base code for each country and overloaded only the specific different behavior of every country, like for the payment methods, for example Brazil supporting “parcelados” or Germany with specific payment players.

We rewrote the old procedural PHP BackEnd into modern PHP, with OOP and our own Framework but we had to keep the transactional code in existing MySQL Procedures, so the logic was split. There was a Front End Team consuming our JSONs. Basically all the Front End code was cached in Akamai and pages were rendered accordingly to the JSONs served from out BackEnd.

It was a huge success.

This e-Commerce site had Campaigns that started at a certain time, so the amount of traffic that would come at the same time would be challenging.

The project was working very well, and after some time the original Team was split into different projects in the company and a Team for maintenance and evolutives was hired.

At certain point they started to encounter duplicate transactions, and nobody was able to solve the mystery.

I’m specialized into fixing impossible problems. They used to send me to Impossible Missions, and I am famous for solving impossible problems easily.

So I started the task with a SRE approach.

The System had many components and layers. The problem could be in many places.

I had in my arsenal of tools, Software like mysqldebugger with which I found an unnoticed bug in decimals calculation in the past surprising everybody.

Previous Engineers involved believed the problem was in the Database side. They were having difficulties to identify the issue by the random nature of the repetitions.

Some times the order lines were duplicated, and other times were the payments, which means charging twice to the customer.

Redis Cluster could also play a part on this, as storing the session information and the basket.

But I had to follow the logic sequence of steps.

If transactions from customer were duplicated that mean that in first term those requests have arrived to the System. So that was a good point of start.

With a list of duplicated operations, I checked the Webservers logs.

That was a bit tricky as the Webserver was recording the Ip of the Load Balancer, not the ip of the customer. But we were tracking the sessionid so with that I could track and user request history. A good thing was also that we were using cookies to stick the user to the same Webserver node. That has pros and cons, but in this case I didn’t have to worry about the logs combined of all the Webservers, I could just identify a transaction in one node, and stick into that node’s log.

I was working with SSH and Bash, no log aggregators existing today were available at that time.

So when I started to catch web logs and grep a bit an smile was drawn into my face. :)

There were no transactions repeated by a bad behavior on MySQL Masters, or by BackEnd problems. Actually the HTTP requests were performed twice.

And the explanation to that was much more simple.

Many Windows and Mac User are used to double click in the Desktop to open programs, so when they started to use Internet, they did the same. They double clicked on the Submit button on the forms. Causing two JavaScript requests in parallel.

When I explained it they were really surprised, but then they started to worry about how they could fix that.

Well, there are many ways, like using an UUID in each request and do not accepting two concurrents, but I came with something that we could deploy super fast.

I explained how to change the JavaScript code so the buttons will have no default submit action, and they will trigger a JavaScript method instead, that will set a boolean to True, and also would disable the button so it can not be clicked anymore. Only if the variable was False the submit would be performed. It was almost impossible to get a double click as the JavaScript was so fast disabling the button, that the second click will not trigger anything. But even if that could be possible, only one request would be made, as the variable was set to True on the first click event.

That case was very funny for me, because it was not necessary to go crazy inspecting the different layers of the system. The problem was detected simply with HTTP logs. :)

People often forget to follow the logic steps while many problems are much more simple.

As a curious note, I still see people double clicking on links and buttons on the Web, and some Software not handling it. :)