Open Infrastructure Summit, Denver 2019

The first ever Open Infrastructure Summit was held in the last week of April 2019 at the Colorado Convention Center in Denver, CO. It’s the first since the re-branding from OpenStack to Open Infrastructure began last year to be officially held with the new name. Otherwise, it felt just like the OpenStack summits of old.

The keynotes were better than in prior summits – I think the sponsors got the feedback that no one was interested in sitting through a recap of “how they did X with OpenStack”, and instead focused more on what they intended to do with it. There was a great demo by Chris Hoge and Julia Kreger that showed a kubernetes operator managing a bare metal infrastructure; it showed very clearly that the typical media message around “Kubernetes is replacing OpenStack” is silly. They exist in different problem spaces, and work well together. The only place Kubernetes is replacing OpenStack is in the hype cycle.

After the keynotes I went to the Nova Project Update session. It was very thorough, but felt more like someone reading release notes out loud. I had hoped for more of a discussion about the thinking that went into some of the things that were worked on or are being planned rather than just a straight recitation.

After that was lunch – sort of. For the first time since these summits began, lunch was not provided. Instead, you were supposed to go to one of the many restaurants in the area and buy your own lunch. However, since we had pretty poor weather—freezing temperatures, snow, and rain—walking around downtown Denver wasn’t what I felt like doing. Judging by how packed the restaurant in the hotel across the street was, a lot of other people felt the same way. I understand that times are not as heady as in previous years when OpenStack was the latest hotness, but this seemed like a poor place to cut back. I always enjoyed sharing a table with a bunch of other OpenStackers and learning about where they were from and what they were doing with OpenStack. Going out to lunch meant that people tended to stay with groups they already knew. The afternoon snacks were also gone, which is no big deal for me, but others mentioned to me that they missed having them. Finally, they didn’t have a signature piece of conference swag. I’m typing this wearing the OpenStack hoodie I got back in the Paris 2014 summit, and have my sweatshirt from Tokyo 2015 in my room. Well, OK, they did give out a pair of socks, but they weren’t tied to the event. It’s not a huge thing, but not having something this time really makes things feel… different. And not in a good way.

There weren’t any sessions in the afternoon that I really wanted to go to, so instead I worked on two OpenStack-related projects: etcd-compute and using Graph Databases, such as Neo4j, to hold information for the Placement service. I have previously written about my work with both of these. And since the author of etcd-compute, Chris Dent, was also here at the summit, it was a perfect time to work on it together, so I set up several VMs for us to “play with”.

Monday evening after the sessions was the “Marketplace Mixer”, which is a way to get the attendees to visit the vendor area. They provided food and beverages, and I had my badge scanned several times in exchange for some local craft beer. There wasn’t a lot offered by the vendors that would be useful to me, but I did run into a lot of people I knew. When you’re in your 10th year of working on OpenStack, you get to know quite a few people!

On Tuesday I started with a session on Nova-Cyborg integration. Or at least that was what it was advertised as. It turned out to be more of an “Introduction to Cyborg Concepts” talk, rather than focusing on where the two projects needed to integrate.

The crowd at the Cyborg-Nova integration session

Later on was the API-SIG BoF (Birds of a Feather) session that I headed up. There hadn’t been much traffic in the SIG ahead of the summit, so I was happily surprised when several people showed up. We ended up having a good discussion on a variety of API-related topics, and I got to meet several of the people who have joined in some of the more recent IRC discussions and Office Hours who previously I had only known by their IRC handles. It’s always nice to put a face to a name.

In the afternoon was a session to update everyone on the process of extracting Placement from Nova. In the past this has been a somewhat heated topic, but this time everyone seemed to understand where things were and were pretty cool with it. There weren’t any long discussions, so the session finished early. I guess that’s a very good sign that we handed that process well.

The final session of the afternoon was to discuss what the various SIGs (Special Interest Groups) and WGs (Working Groups) needed to be successful. Since the API-SIG has been around for many years, we didn’t really have any needs along these lines. Sure, it would be great to get more people involved, but it isn’t critical. Some of the newer groups explored ways of getting the word out about their existence, which is always a problem. There is so much going on in the OpenStack world that getting people to pay attention to yet another thing is always challenging.

That evening was the Open Infrastructure party, sponsored by Trilio, Mirantis, Red Hat, Open Telekom Cloud, & AVI Networks. It was held in The Church Nightclub, which is an old church that has been converted to a nightclub. There was an open bar and food available, and they had a band playing for entertainment. The location was fun, but being indoors with loud music meant that there was only so much conversation you could have. Still, it was fun!

Open Infrastructure Party
The crowd at the Open Infrastructure Party at the Church Niteclub
church niteclub
A view from higher up that shows how an old church was converted into a niteclub. You can see the some of the band playing at the very bottom.

There weren’t any talks on Wednesday morning that I really wanted to attend, so I spent most of the morning in the designated hacking room working on the etcd-compute project for a while, and then on implementing many of the features that are currently lacking in Placement in my graph database code. I managed to implement passing a tree structure to represent nested resource providers so that it creates the corresponding nodes and relationships in the database. This implementation is becoming more and more complete, and I hope when I show it to others this week that they are able to get out of their MySQL comfort zone and see how much better this approach is for representing resources.

I went to lunch with some of the members of my team at IBM who were at the Summit, along with some people from Red Hat with whom we are working to ensure that their various offerings run as well on Power hardware as on x86. So while the pizza was tasty, it was definitely a working lunch. It was also great to meet some of the people I had only known online before.

The Red Hat – IBM lunch *after* the food had been eaten.

After lunch was a session focused on the gaps between Nova functionality and what has been implemented in OpenStack Client. Most of the missing functionality is concerned with supporting new microversions, and this support is several years behind. I’m not sure how effective the discussions were, since what is really needed is for people to take ownership of some of the needed tasks, and I didn’t hear a lot of that happening.

After that I went to the Cyborg Project Update. Once again, it probably would have been much more useful to anyone who hadn’t been following along with the project, so while I didn’t get much from it, there was a lot of information presented on the current state and future plans for Cyborg.

And that was it! The end of another Summit, even if it was the first. That evening I met my sister for dinner. She lives in the Denver area, and it was great to catch up with her and spend some time relaxing after 3 long days. But the relaxation will be short-lived, as the Train PTG starts first thing tomorrow morning!

Geri & Ed
Selfie with my big sister Geri

More fun with etcd-compute

Last time I ended my work getting etcd-compute running at the point where I needed to configure the virtual networking. I’ve been busy the past few days with meetings and other work-related stuff, so it’s taken me a while to continue on this experiment. But I have some time now; let’s jump back in!

The reason I thought that I needed to set up virtual networking was that when I ran ip a on my controller node, all I had was the loopback and main ethernet interfaces. The directions for etcd-compute talked about setting up the metadata server by adding the IP address it uses to a virtual bridge: sudo ip addr add dev virbr0. As I didn’t have such a bridge on my VM, I figured I had to add it. I tried sever guides on adding a bridge to an Ubuntu server, but each one ended up messing up the networking, making the VM unreachable. I ended up re-creating my etcd1 so many times that I gave up and figured I try without the metadata server. I started the placement and etcd servers by running, and then just on a lark I re-ran ip a. This time it showed:

ed@etcd1:~$ ip a
1: lo: mtu 65536 qdisc noqueue state UNKNOWN group default qlen 1000
link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
inet scope host lo
valid_lft forever preferred_lft forever
inet6 ::1/128 scope host
valid_lft forever preferred_lft forever
2: ens3: mtu 1500 qdisc fq_codel state UP group default qlen 1000
link/ether fa:16:3e:90:6d:d0 brd ff:ff:ff:ff:ff:ff
inet brd scope global ens3
valid_lft forever preferred_lft forever
inet6 fe80::f816:3eff:fe90:6dd0/64 scope link
valid_lft forever preferred_lft forever
3: virbr0: mtu 1500 qdisc noqueue state DOWN group default qlen 1000
link/ether 52:54:00:35:c1:0d brd ff:ff:ff:ff:ff:ff
inet brd scope global virbr0
valid_lft forever preferred_lft forever
4: virbr0-nic: mtu 1500 qdisc fq_codel master virbr0 state DOWN group default qlen 1000
link/ether 52:54:00:35:c1:0d brd ff:ff:ff:ff:ff:ff

I’m not sure how those entries for ‘virbr0’ and ‘virbr0-nic’ got added (maybe docker added them?), but I wasn’t going to worry about that! So I ran the following commands, and they worked without a problem:

sudo ip addr add dev virbr
sudo python md_server/mdserver/ mdserver.conf &

So now that the metadata server is running, time to try running ecompute on all the nodes. I use iTerm2, which has some sweet tools for splitting the terminal screen and running the same command in the different panes. I recorded a script of what happened:

I ran the command ecompute & on all the nodes to start the compute service in the background.

ed@etcd1:~/projects/etcd-compute(master)$ ecompute &
[1] 4661
ed@etcd1:~/projects/etcd-compute(master)$ 1556230694.3633301: PID: 4661 [None] {'uuid': '19a89e30-4bdd-49e7-b1a0-d4172bf7b289', 'placement': {'endpoint': 'http://etcd1:8080'}, 'etcd': {'host': 'etcd1'}, 'resize': False, 'bridge': 'br0'}
1556230694.364856: PID: 4661 [19a89e30-4bdd-49e7-b1a0-d4172bf7b289] {'VCPU': 4, 'DISK_GB': 77, 'MEMORY_MB': 7976}
1556230694.5012665: PID: 4661 [19a89e30-4bdd-49e7-b1a0-d4172bf7b289] Existing resource provider with gen 7 found with usages: VCPU: 0, MEMORY_MB: 0, DISK_GB: 0.

It’s interesting to see that because I had run this a few times earlier, etcd-compute recognized the UUID of the node, and noted that there was already an entry for that resource provider, with a generation of 7. If I were to stop that ecompute service and then re-start it, I would see the same as above, except this time the generation would be 8. That’s because when the service is killed, it changes the ‘reserved’ amount of its VCPU inventrory to the total amount, effectively preventing that node from being provisioned. That change increments the resource provider’s generation.

At about the 30-second mark, I tried to create a VM by running the command eschedule 'resources=VCPU:1,DISK_GB:1,MEMORY_MB:256' on the etcd3 node. That worked, and almost immediately you can see that it was scheduled to the etcd1 node, and the build process starts. However, there were many errors output, with the main one being error: failed to get domain ‘ff77fe58-e96a-498b-a3f5-a59030987238’. This is repeated several times, along with a bunch of network errors. So at this point I stopped the experiment.

There’s a lot I learned by going through all this, and I see many places where the etcd-compute project could be improved, starting with the documentation. I’d also like to get some less ethereal debugging output, so that when there are problems like I had spinning up a VM, they are recorded for later analysis. I’d also like to learn a lot more about the details of the networking required so that I can make sense of some of the networking errors.

The author of etcd-compute, Chris Dent, and I are hoping to have a mini-sprint on this project next week at the Open Infrastructure Summit in Denver, Colorado. If you will be there and want to join in the fun, drop me an email and I’ll let you know when we settle on a time and place.

Playing with etcd-compute

I’ve been interested in the etcd-compute project by Chris Dent. It’s sort of a lightweight virtual machine manager like OpenStack Nova, but without the complexity and cruft Nova has accumulated over the past 9 years. It takes advantage of technologies that simply didn’t exist in 2010 when Nova was created, using etcd‘s built-in notifications instead of passing large, complex objects over a message bus to make Remote Procedure Calls (RPC).

Keep in mind that Nova does a lot of things that etcd-compute can’t, so this isn’t a potential 1:1 replacement for Nova. But it does have potential as a much lighter replacement for those applications where the full power of Nova isn’t needed.

This post is designed to be obsolete within a week or so. What I’m aiming for is to record what worked for me following Chris’s instructions. Where I run into problems shows one of three things: our systems start out differently, or Chris assumed something that wasn’t in the file, or my brain is not firing on all cylinders. It is my hope that this may help improve the installation instructions, and guide others who may wish to explore etcd-compute.

I don’t have a lot of hardware—ok, any hardware—at my disposal to experiment with, so I started by creating 3 Ubuntu 18.04 VMs in the internal OpenStack cloud for my team here at IBM. Yes, you can run virtualization on top of virtualization, and it’s turtles all the way down. But it does work! I named the instances etcd1, etcd2, and etcd3, with etcd1 being the controller and the others used as standard compute nodes.

There are some requirements—, virtinst, libvirt-daemon, libvirt-clients, and libguestfs-tools—that need to be installed on all the nodes, so I updated the distro packages and installed the requirements. Unfortunately, libvirtd wouldn’t start, and well, that’s kind of an important piece. So I cleaned house and tried again:

ed@etcd1:~$sudo aptitude purge libvirt-daemon
ed@etcd1:~$sudo apt install -y qemu qemu-kvm libvirt-bin  bridge-utils  virt-manager
ed@etcd1:~$ sudo systemctl enable libvirtd.service
Synchronizing state of libvirtd.service with SysV service script with /lib/systemd/systemd-sysv-install.
Executing: /lib/systemd/systemd-sysv-install enable libvirtd
Created symlink /etc/systemd/system/libvirt-bin.service → /lib/systemd/system/libvirtd.service.
Created symlink /etc/systemd/system/ → /lib/systemd/system/virtlockd.socket.
Created symlink /etc/systemd/system/ → /lib/systemd/system/virtlogd.socket.
ed@etcd1:~$ sudo systemctl start libvirtd.service
ed@etcd1:~$ sudo systemctl status libvirtd.service
● libvirtd.service - Virtualization daemon
Loaded: loaded (/lib/systemd/system/libvirtd.service; enabled; vendor preset: enabled)
Active: active (running) since Tue 2019-04-23 23:29:30 UTC; 5s ago
Docs: man:libvirtd(8)
Main PID: 5289 (libvirtd)
Tasks: 19 (limit: 32768)
CGroup: /system.slice/libvirtd.service
├─4486 /usr/sbin/dnsmasq --conf-file=/var/lib/libvirt/dnsmasq/default.conf --leasefile-ro --dhcp-script=/usr/lib/libvirt/libvirt_
├─4487 /usr/sbin/dnsmasq --conf-file=/var/lib/libvirt/dnsmasq/default.conf --leasefile-ro --dhcp-script=/usr/lib/libvirt/libvirt_
└─5289 /usr/sbin/libvirtd
Apr 23 23:29:30 egleafe-etcdcompute-1 systemd[1]: Starting Virtualization daemon…
Apr 23 23:29:30 egleafe-etcdcompute-1 systemd[1]: Started Virtualization daemon.
Apr 23 23:29:30 egleafe-etcdcompute-1 dnsmasq[4486]: read /etc/hosts - 10 addresses
Apr 23 23:29:30 egleafe-etcdcompute-1 dnsmasq[4486]: read /var/lib/libvirt/dnsmasq/default.addnhosts - 0 addresses
Apr 23 23:29:30 egleafe-etcdcompute-1 dnsmasq-dhcp[4486]: read /var/lib/libvirt/dnsmasq/default.hostsfile
lines 1-17/17 (END)

So I guess it’s working now!

It’s also a pain to always have to use sudo to run docker commands, so add your user to the docker group. The command for this is sudo usermod -a -G docker ed, which adds user ‘ed’ to the group ‘docker’. You have to log out and log back in for it to take effect, but once you do, you can run commands like docker ps -a without sudo.

Also, in my experience I’ve run into various odd problems using the distro version of Python, so I prefer to install from source to get the latest Python (3.7.3 right now).

Being a creature of habit, I like having the project code I’m working with to be under a ~/projects directory. So for each of these instances, I ran the following:

ed@etcd1:~$ mkdir projects
ed@etcd1:~$ cd projects/
ed@etcd1:~/projects$ git clone
Cloning into 'etcd-compute'…
remote: Enumerating objects: 169, done.
remote: Counting objects: 100% (169/169), done.
remote: Compressing objects: 100% (107/107), done.
remote: Total 205 (delta 97), reused 124 (delta 61), pack-reused 36
Receiving objects: 100% (205/205), 54.58 KiB | 119.00 KiB/s, done.
Resolving deltas: 100% (107/107), done.
ed@etcd1:~/projects$ cd etcd-compute/

As the etcd-compute code has its own dependencies, those need to be installed by running sudo python develop. When I ran that the first time, I got an error when it was trying to install libvirt-python. I tried installing some other libvirt-related libraries and binaries, but I kept getting the same error. After a while I was trying anything I could think of, even running under Python 2! (didn’t work). Maybe it was something about Python 3.7 that was problematic, so I created a venv for Python 3.6, and ran pip install libvirt-python. It installed without a problem. Hmmm. So I fired up a Python 3.7 venv, and it also installed into that. It seems that the installation using was doing something different than a straight pip install. To test that, I got rid of the venvs, and ran sudo pip install libvirt-python, and it worked just fine. I was then able to install the rest of the dependencies by running sudo python develop.

Now that the dependencies are installed, we need to create the database for placement, and then modify the dockerenv file so that the OS_PLACEMENT_DATABASE__CONNECTION setting points to that. My database is on a MariaDB server, so I needed to change the value to:


That means, of course, that I need to install pymysql using sudo pip install pymysql before I can make a connection. Once that’s done, I started the docker containers by running ./ from the primary VM. In my case, that’s etcd1.

That brings up another edit that’s needed on all your “machines”: changing the location of the host in compute.yaml and schedule.yaml. These assume that the host is named ‘ds1’, which isn’t true in this case. I changed ‘ds1’ to ‘etcd1’, and then added an entry in each node’s /etc/hosts file with the IP address of the etcd1 VM.

We also want to create a value for the uuid in compute.yaml. One simple way is to run python -c "import uuid; print(uuid.uuid4())", and copy the output to paste into compute.yaml. Do that on every compute node you are running.

That’s enough for one day. Tomorrow we start with configuring networking!

Using a Python virtual environment

This is a quick demonstration of how to create a virtual environment in Python3. I’m starting with an empty directory in ~/projects/demo. I then run the command to create a virtual environment:

ed@imac:~/projects/demo$ ll
ed@imac:~/projects/demo$ python3 -m venv my_env
ed@imac:~/projects/demo$ ll
total 0
drwxr-xr-x 6 ed staff 192B Jan 24 18:14 my_env

Note that the command created a directory with the name I gave it: ‘my_env’. Next we have to activate it. ‘Activate’ changes Python’s internal references to look for things such as which Python version to run, and where installed modules are placed.

ed@imac:~/projects/demo$ source my_env/bin/activate (my_env)ed@imac:~/projects/demo$

I have a bash script that changes the prompt to show the current Python environment; notice that after activating the prompt now starts with ‘(myenv)’.

Installed modules are located in the ‘site-packages’ subdirectory that’s a few levels deep. Let’s see what’s in this fresh virtual env’s site-packages:

(my_env)ed@imac:~/projects/demo$ ll my_env/lib/python3.6/site-packages/
total 8
drwxr-xr-x 3 ed staff 96B Jan 24 18:14 pycache
-rw-r--r-- 1 ed staff 126B Jan 24 18:14
drwxr-xr-x 23 ed staff 736B Jan 24 18:14 pip
drwxr-xr-x 10 ed staff 320B Jan 24 18:14 pip-9.0.1.dist-info
drwxr-xr-x 6 ed staff 192B Jan 24 18:14 pkg_resources
drwxr-xr-x 34 ed staff 1.1K Jan 24 18:14 setuptools
drwxr-xr-x 12 ed staff 384B Jan 24 18:14 setuptools-28.8.0.dist-info

One of my favorite development tools for Python is the pudb debugger. To show that we can install a package, let’s try importing it first (and failing):

(my_env)ed@family-imac:~/projects/demo$ python
Python 3.6.4 (v3.6.4:d48ecebad5, Dec 18 2017, 21:07:28)
[GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import pudb
Traceback (most recent call last):
File "", line 1, in
ModuleNotFoundError: No module named 'pudb'
>>> ^D

Now let’s install it using pip:

(my_env)ed@family-imac:~/projects/demo$ pip install pudb
Collecting pudb
Collecting urwid>=1.1.1 (from pudb)
Collecting pygments>=1.0 (from pudb)
Using cached
Installing collected packages: urwid, pygments, pudb
Successfully installed pudb-2018.1 pygments-2.3.1 urwid-2.0.1

Let’s look at the site-packages directory after installing pudb:

(my_env)ed@family-imac:~/projects/demo$ ll my_env/lib/python3.6/site-packages/
total 8
drwxr-xr-x 10 ed staff 320B Jan 24 18:38 Pygments-2.3.1.dist-info
drwxr-xr-x 3 ed staff 96B Jan 24 18:14 pycache
-rw-r--r-- 1 ed staff 126B Jan 24 18:14
drwxr-xr-x 7 ed staff 224B Jan 24 18:35 pip
drwxr-xr-x 9 ed staff 288B Jan 24 18:35 pip-19.0.1.dist-info
drwxr-xr-x 6 ed staff 192B Jan 24 18:14 pkg_resources
drwxr-xr-x 18 ed staff 576B Jan 24 18:38 pudb
drwxr-xr-x 9 ed staff 288B Jan 24 18:38 pudb-2018.1.dist-info
drwxr-xr-x 22 ed staff 704B Jan 24 18:38 pygments
drwxr-xr-x 34 ed staff 1.1K Jan 24 18:14 setuptools
drwxr-xr-x 12 ed staff 384B Jan 24 18:14 setuptools-28.8.0.dist-info
drwxr-xr-x 33 ed staff 1.0K Jan 24 18:38 urwid
drwxr-xr-x 7 ed staff 224B Jan 24 18:38 urwid-2.0.1.dist-info
(my_env)ed@family-imac:~/projects/demo$ python

Note that there are now entries for both pudb and its dependency, pygments. And to verify that it has been successfully installed:

(my_env)ed@family-imac:~/projects/demo$ python
Python 3.6.4 (v3.6.4:d48ecebad5, Dec 18 2017, 21:07:28)
[GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import pudb
>>> pudb.version
>>> ^D

There’s a ton more to using virtual environments, but that should give you a start.

Graph Database Follow-up

I just published my series on Graph Databases a few hours ago, and have already had lots of hits. I guess people found the topic as interesting as I did! One reader, Jay Pipes, is one of the core developers of the Placement service I mentioned in that series, and probably the most heavily opinionated with regards to using MySQL. He is also a long-time friend. Jay asked some excellent questions at the end of the last post, and I didn’t feel that I could do them justice in the space of a WordPress reply box, so I created this post. Let’s get into it!

1) I note that you’ve combined allocations and inventory into a single object. How would we query for the individual consumers of resources?

The combination was for simplifying the illustration. As you can probably guess, a full Placement model would have Consumer objects that would have a [:CONSUMES] relation to the resource objects. So to illustrate this, I added another script to the GitHub repo: Since  relations can have properties too, we’d create an ‘amount’ property that would specify how much of a resource is being consumed by this particular consumption. The query would look something like this:

MATCH (disk:DISK_GB)<-[alloc:CONSUMES]-(con:Consumer)
WITH disk, - SUM(alloc.amount) AS avail
WHERE avail > 2000
RETURN disk, avail

This would return the disks whose total amount minus the sum of all the consumed amounts, is greater than 2000.

2) How many resources of each type are being consumed by a project or user or consumer? (i.e. usages, aggregated by a grouping mechanism – the thing that SQL does pretty well)

In this respect, it wouldn’t be all that different. Cypher does aggregates, too. Using the data generated by the script, I can run the following query:

MATCH (con:Consumer)-[alloc:CONSUMES]->(r)
RETURN labels(r)[0] AS resource_class, sum(alloc.amount) AS usage

This query will return:

│"resource_class"   │"usage"                │
│"MEMORY_MB"        │1024                   │
│"DISK_GB"          │500                    │
│"VCPU"             │2                      │

So as you can see, not a whole lot different than with SQL.

3) How do you ensure that multiple concurrent writers do not over-consume resources on a set of providers? We use a generation on the provider along with an atomic update-where-original-read-generation strategy to protect the placement DB objects. I’m curious how neo4j would allow that kind of operation.

Neo4j is ACID-compliant, and transactions either succeed completely or not at all. So I’m not really sure how much I can add to that. Setting the value of an object is trivial, updating that value is trivial, checking a supplied value with the one stored in the DB is trivial. So it would work identically to how it works in SQL.

In summary, I want to make two points: first, I’m not bashing MySQL. It’s an excellent database that I use extensively myself (well, I use MariaDB, but…). This WordPress site has a MariaDB backend. Second, this started as an attempt to scratch the itch I have felt since we began making Placement more and more complex. I wanted to find out if there was a better way to handle the sort of problems we have, and my previous reading on graph databases came to mind. It was tough to get used to thinking with a graph mindset. It reminded me of when I first started using Go after years of Python: when I tried to write Pythonic code in Go, it was terrible. Once I was able to drop the Python mindset and write Go the way it was designed, things were much, much easier to understand, and the code I wrote was much, much better. The same sort of “letting go” process had to happen before I was able to fully see what can be done with graph databases.