Day 87 of DevOps, Day 37 of Azure. Both tasks today turned on a name that is not the name the system uses.
An Ansible task that says yum runs the dnf module, because the first is now a redirect to the second. And an Azure image "published by Jetware" is published by jetware-srl, which no amount of guessing gets you to.
One Ansible task, one Azure task. Install and start a service on three servers, then put MySQL on a VM and prove a PHP page can reach it. The tasks come from the KodeKloud Engineer platform.
Task 1: Install vsftpd and start it
The task: a playbook that installs vsftpd on all three app servers and starts it.
Step 1: Check it isn't there
[tony@stapp01 ~]$ sudo systemctl status vsftpd Unit vsftpd.service could not be found. Same discipline as the last batch, and it is what makes changed: [stapp01] mean "installed" rather than "already fine".
Step 2: The playbook
--- - hosts: all become: yes tasks: - name: Install vsftpd package. yum: name: vsftpd state: present - name: Starting the vsftpd service. service: name: vsftpd state: started enabled: yes stapp01 : ok=3 changed=2 stapp02 : ok=3 changed=2 stapp03 : ok=3 changed=2 Both modules are idempotent, so a second run reports changed=0. Which, after Day 84's state: touch reporting a change forever, is worth saying out loud: package and service describe a desired state, and a system already in that state needs no action.
enabled: yes alongside state: started is two separate claims, and the verification shows both:
Loaded: loaded (/usr/lib/systemd/system/vsftpd.service; enabled; preset: d> Active: active (running) since Sat 2026-05-23 14:43:50 UTC; 11s ago enabled in the first line means it comes back after a reboot. active (running) in the second means it is up now. Day 23 made the same point about systemctl enable --now.
Step 3: The task says yum and the module is dnf
ansible.builtin.yum's documentation page now says one thing: "This is a redirect to the ansible.builtin.dnf module."
So what executes is dnf. The task name says yum, the playbook says yum, and the module doing the work is the dnf module. Which is correct for these servers, because Day 71 established the app servers are CentOS Stream 9, where dnf is the package manager and yum is a compatibility name.
Three things follow, and they pull in different directions.
The playbook is right as written. The redirect handles it, and nothing needs changing.
It is misleading to read. Anyone debugging this is reading dnf's documentation whether they know it or not, and the two modules' parameters have not always matched.
And it is not portable. Both yum and dnf commit the playbook to the RHEL family. ansible.builtin.package dispatches to whatever the target actually uses, so the same playbook runs on Debian hosts too.
What the docs recommend
For a lab that names yum, yum is the right answer. For anything meant to outlive one distribution, package is the module that does not need rewriting later.
One other thing worth recording, and it is mildly annoying. The inventory for this task has ansible_ssh_pass back in it, for all three hosts, in plain text. Day 82's whole task was removing exactly those, and Ansible's documentation says of ansible_password, "never store this variable in plain text; always use a vault."
Each lab starts fresh, so the plain-text passwords come back with it. The fix has to be redone every time, which is a reasonable description of security work generally.
Task 2: MySQL on one VM, PHP on another
The task: a VM from a marketplace Percona image in Central US, a PHP VM in East US, and a page that connects across.
Step 1: The publisher is not the brand
The task said "published by Jetware".
az vm image list --all --publisher jetware --location centralus (empty) What I found was searching on a fragment rather than guessing the whole thing:
az vm image list-publishers --location centralus --query "[?contains(name,'etware')].name" -o tsv etware, without the leading letter, because marketplace publisher names are inconsistently capitalised and both spellings are plausible. The answer is jetware-srl.
Image identity is a four-part URN, and none of the parts are derivable from the product name:
jetware-srl : percona_mysql : percona_mysql57-ubuntu-1604 : 1.0.170503 publisher : offer : sku : version Day 3 found the same for Canonical's Ubuntu offers, where the lesson was that offer names cannot be guessed. Here the publisher cannot either.
Which is this morning's Ansible finding wearing different clothes. A module called yum is a redirect to one called dnf. A product called Jetware is published by jetware-srl. Both times the name you are handed is not the name the system uses, and both times the cost of assuming is one lookup you did not do.
Worth noting the image is from December 2017, on Ubuntu 16.04, with Percona 5.7.17. Marketplace images are frozen at publication, and this one has had no updates in nine years. Fine for a lab, a supply-chain problem anywhere else.
Step 2: A denial that told me nothing
az vm image terms accept --urn $URN (AuthorizationFailed) ... does not have authorization to perform action '...agreements/read' Marketplace images carry legal terms that must be accepted per subscription before deployment, and the lab identity cannot read or accept them.
The tempting conclusion is that the image cannot be deployed. That does not follow. The denial is on reading the agreement, which says nothing about whether the agreement is already accepted. My permission to query a setting and the setting's value are independent facts.
So I tried the create anyway, and it worked. The terms were already accepted for the subscription.
This is the mirror of my Day 31 mistake, where a quota error left its Location field blank and I concluded the limit was subscription-wide. Same error, opposite direction: an error about your ability to observe something is not evidence about the thing itself. Attempt the operation when it fails safely.
Step 3: Three gates, checked separately
Password authentication means no SSH key, so every in-guest operation went through Run Command, which runs as root and returns synchronously.
A remote MySQL connection has three gates, and all three failures look alike from the PHP side:
| Gate | Check | Result |
|---|---|---|
| Azure network | az network nsg rule list | open-port-3306 at 1010 |
| MySQL user host | SELECT user,host FROM mysql.user | xfusion_user / % |
| MySQL bind address | SHOW VARIABLES LIKE 'bind_address' | * |
The bind address is the one most worth checking, because MySQL defaults to 127.0.0.1 in many distributions, and a localhost-only bind defeats a correct NSG rule and a correct % host with nothing to indicate why. This image binds to *.
Checking all three before touching the PHP file meant a later failure could only be the PHP file.
What the docs recommend
Two techniques from this task I would reuse, and one thing the task asks for that production should not.
Base64 for awkward payloads. The SQL contains nested quotes and has to survive the local shell, the Run Command transport and the remote shell:
export SQL_B64=$(printf '%s\n' "CREATE USER ... '%' ..." | base64 -w 0) az vm run-command invoke ... --scripts "echo $SQL_B64 | base64 -d > /tmp/s.sql; /jet/enter mysql < /tmp/s.sql" Base64 output is only A-Za-z0-9+/=, none of which any shell reinterprets. The quoting problem is sidestepped rather than solved, which is better.
And probing an unfamiliar wrapper before scripting against it. /jet/enter mysql is a Jetware convention and I had no idea whether it took -e:
ls -la /jet head -20 /jet/enter /jet/enter mysql -e "SELECT VERSION();" Three cheap commands in one call, answering "does this work" and "if not, what next" at the same time. It turned out to be exec /jet/own/bin/run "$@", a thin wrapper that passes everything through.
The thing production would not do is the architecture:
PHP (eastus) --public internet--> MySQL (centralus, port 3306) MySQL's wire protocol is not encrypted by default, so credentials cross the internet in the clear. az vm open-port with no source restriction opens 3306 to 0.0.0.0/0, defended by password123. And every query pays cross-region latency for no benefit.
The real shape is both VMs in peered VNets talking over private IPs, the NSG rule scoped to the peer's address space, TLS on the connection, and credentials from Key Vault instead of literals in a PHP file. Or Azure Database for MySQL with a private endpoint and no VM at all.
What I learned
Two names, neither of which was the name.
yum is the more interesting one because it is invisible. The playbook works, the output is clean, and the module that ran is not the one written in the file. You only find out by reading a documentation page that contains a single sentence.
jetware-srl is the louder version. An empty result from a search is at least a signal that something is wrong, and the fix was searching a fragment instead of a guess.
The one I would take into real work is the denial. An AuthorizationFailed on reading the marketplace agreement looks like a wall, and all it says is that I cannot see the answer. The answer happened to be yes. Twice now in this series I have reasoned from the absence of information as though it were information, and twice the one command I had already written would have settled it.
So here is the Day 87 question. What have you concluded recently from an error that only told you what you were not allowed to look at?
Day 87 down. Thirteen to go.
