Hacker News new | past | comments | ask | show | jobs | submit login
Terraform Gotchas and How We Work Around Them (heap.engineering)
158 points by kalmar on June 12, 2017 | hide | past | web | favorite | 80 comments

> Always write your plan -out, and apply that plan

I have in my dotfiles:

    alias tfplan='terraform plan -out=.tfplan -refresh=false'
    alias tffreshplan='terraform plan -out=.tfplan'
    alias tfapply='terraform apply .tfplan; rm .tfplan'
That way I never accidentally `terraform apply` without creating a plan first. I also have it not refresh the state by default, which is mostly unnecessary and speeds up the planning significantly.

Hey all - Seth here from HashiCorp (the company that makes Terraform). The next version of Terraform (0.10) natively adopts very similar behavior, presenting a plan before applying as an added safety step. You can read more in the 0.10 upgrade guide. At the time of this writing, 0.10 is not yet released, but compiling Terraform from source at master will inherit this behavior.


I feel bad about not following up on it yet but my comment on TF#13276 sums up the issues I have with Terraform after using it for a little under a year now.


I hope you all can work on improving the definitions, because many of them really are a chore compared to setting things up in the AWS dashboard, at the moment (security groups for example).

Oh interesting. Note to self: see if there's an option to disable `terraform apply` without a plan.

I always refresh when running the pre-apply plan, but while iterating I use that. Do you always run your `tffreshplan` command before applying?

Agree would be nice to add a new option which prevents invoking `apply` without a supplied plan argument. Create an issue in the GitHub, I'll upvote.

My muscle memory for "terraform apply" was too strong, so I just moved "terraform" out of my path altogether. Now I have a wrapper that invokes it safely (or a Makefile, depending on which project I'm working on).

Terraform has interested me for a while, and I've been meaning to give it a try, but haven't had a chance just yet.

From what I have seen so far though, there isn't really that much difference/benefit over CloudFormation. We currently have 95% of our resources in AWS with about 4% in Azure, and 1% in Google Cloud. It's great that Terraform is 'mulit-cloud' but it still seems like you have to write .tf's catered to each cloud, you can't just lift and shift to another cloud by copying and pasting a file?

People say the 'plan' feature is one of the advantages over CFN, but as far as I can tell, CFN now offers the same feature... it tells you what's going to change when you upload a new stack.

I sound like a CFN advocate now, but I genuinely don't have that much experience with it, and really do want to give Terraform a chance. Convince me?

Oh, and since CFN started supporing YAML it looks easier to write too

I've been thru the CFN v TF question. We came up with a list of benefits of TF over CFN. (Yes, I know - one-sided, but we wanted to document the decision with a bit more substance than "oh it's just better")

  * Ability to separate data (variables/parameters) from configs.
  * Easier to read (well at least pre-YAML CFN). 
  * Allows comments in the code.
  * Version control changes (diffs) are easier to read.
  * Multi-Cloud support. Works against AWS, Google Compute, Azure, Docker, more.
  * Multi-provider in general: can provision resources across multiple different cloud providers at once.
  * Can write modules in TF that can be reused in multiple different configs.
  * Tracks state via a version-controllable data file.
  * 'terraform plan' is essentially a no-op mode to see what changes would occur without actual running or making changes.
  * Actively developed.

Cloud Formation has good support for use from Python, Ruby, Node and the JVM (with template generators, to help out). If you're writing JSON directly, yes, some of the points above -- the first four, and the seventh -- are an issue; but if you use Python you get all the benefits of it being "real code" and "just a library" (unlike Terraform).

This is a situation where I think not being "real code" is a feature of terraform. You declaratively represent your infrastructure rather than generate it with real code.

Over time, I have come to view "declarative infrastructure" as unrealistic. It's right 90% of the time, but not 100% of the time -- kind of like using only CSS and HTML. One should use markup whenever possible; but not everything on a page is truly "declarative". Occasionally one needs to script an input field or a transition.

One example of this is scripting the handoff process that's part of a blue/green deploy. In practice you'll want to look at organization defined metrics. There are libraries to do this -- either internal to your organization, or provided by a metrics vendor -- and scripting the process looks like this:

    (1) Setup new environment.
    (2) Divert some traffic.
    (3) Check metrics.
    (4) If metrics are okay:
        (4.1) Post message internally (IRC/Slack).
        (4.2) Divert all traffic.
        (4.3) Set up timed task to tear down old environment (in a day, hour, &c.).
    (5) If metrics are okay:
        (5.1) Post message internally (IRC/Slack), maybe to different people.
        (5.2) Stop diverting traffic.
        (5.3) Tear down new environment.
A large part of the work here is declarative: (1) by itself is a big piece of it, and is fully declarative, as is the teardown in (4.3) and (5.3). However, the need for control flow in this and many other cases means that, without a library, one must drive Terraform by templating and shelling out. Not being "real code" pushes one in the direction, not of greater declarativeness (libraries can certainly have declarative interfaces, like Troposphere does), but of worse code.

Many complex and powerful features are exposed to a modern business through libraries -- AI, payments, telephony -- and software defined infrastructure can be, too. The benefits of "infrastructure as code" won't be realized until that happens.

A note on the Multi-Cloud support; OP is correct, it's not "switch to another provider" type setup - that's not where the power lies, the cool stuff is about being able to share attributes across clouds - Think about creating an AWS ELB, and then adding that CNAME entry into your CloudFlare account.

It also makes it for when/if you want to switch or start supporting other providers - your tool is already agnostic, and you don't need to go from CloudFormation and port over - you're already there.

I have seen similar comments about "what happens if you want to switch providers" or "don't want to put all your eggs in one basket'. I can understand the comments but at the same time I do a lot of enterprise migrations with companies. We do not go in there while working through a plan on how to migrate to AWS or Google cloud and start promoting the multi provider feature of terraform. Why would you be going through so much effort to plan and move an entire enterprise into the cloud and use terraform and then throw in the idea that "Hey, you can also write the code to stand up similar services in another provider?" Maybe for smaller projects it would be attractive but for the big boys I don't see that to be very beneficial. At least this ability would not sway me to use Terra over CFT.

Here's a real-world benefit I had over the weekend. I've been using Terraform for a Google Cloud project, and was coming to setting up DNS records. I wanted to have dual DNS providers for redundancy. With Terraform, it was trivial, as I just added a second Terraform resource for DNSimple (with Google Cloud DNS as my 'main' DNS resource).

The beauty of Terraform is that you can orchestrate all of your infrastructure, not just the stuff in one stack. If you've got 95%/4%/1% in AWS/Azure/GCP, how do you manage/reference the non-AWS resources? Terraform gives you a unified way to reference and link cross-infrastructure resources.

> It's great that Terraform is 'mulit-cloud' but it still seems like you have to write .tf's catered to each cloud, you can't just lift and shift to another cloud by copying and pasting a file?

We're not quite at the point where even the "comparable" cloud services across clouds are drop-in compatible with each other, so this is not going to be possible for a while for reasons not related to Terraform.

> I sound like a CFN advocate now, but I genuinely don't have that much experience with it, and really do want to give Terraform a chance. Convince me?

You could spend your time learning either a vendor-specific tool (CFN), or a vendor agnostic one (Terraform). Since Terraform can do a lot of what CFN does, it may make sense to spend your time learning Terraform instead.

Edit: not sure about CFN, but Terraform is open source: https://github.com/hashicorp/terraform

I think this sort of abstraction risks making some things awkward and will cater to the lowest common denominator. Better to just be explicit and have a config for each provider.

CloudFormation's approach for post-provisioning deployments is much more limited than the options that Terraform provides as first-class citizens. For example, there's a Chef provisioner as well as local-exec and remote-exec. With CloudFormation, the only options exist in the form of cloud-init that uses userdata.

However, this isn't to say that Terraform is always better than CloudFormation. In fact, I'd prefer to use CloudFormation for A/B | Blue-Green deploys because it supports UpdatePolicy options for rotating newly configured images into an autoscaling group. The logic to do that in Terraform is really not trivial at all (there is no clean, straightforward way to do such deployments with Terraform). Furthermore, rollbacks are significantly more reliable in my experience using CloudFormation than Terraform. Rollbacks may be easier to orchestrate using Terraform than CloudFormation though due to easier reference to non-AWS resources.

My preferred style of AWS deployments and infrastructure-as-code layering is Terraform with broken out modules and generating smaller CloudFormation templates for individual application components that need to be deployed often. This seems like a worst-of-both-worlds option but I think keeping CloudFormation templates constrained to just ASG modifications avoids a lot of the problems.

CloudFormation has a lot of things going for it, but it was unsuitable for us ~18 months ago when I was looking around for a way to "infrastructure as code"-ify our setup, because it can't "adopt" existing resources into your setup, and we have tons of stuff we don't want to bring down/up just to bring under cloudformation.

I used the terraforming gem (either because "terraform import" didnt' exist at the time, or because I just didn't like the way it worked) to bring a lot of things under management by terraform.

Isn't Cloudformation limited to 100 items?

That's not even enough for our Network ACLs.

How do you work around that?

it's 1000 by default now (per account per region) and you can get it raised substantially higher

Really?!? That's awesome.

Docs say 200 though http://docs.aws.amazon.com/AWSCloudFormation/latest/UserGuid...

I've wanted, and tried and failed, to adopt Terraform several times now. What always gets in my way is that we already have all our infrastructure in place, and Terraform's import capabilities are too limited.

For example, the last time I used it, a few months ago, it was not able to import almost any of our Google Cloud stuff, and I discovered that import support is only provided for some resources. There's a third-party tool called Terraforming, but it apparently only works with AWS.

I'm quite disheartened that the world is lagging this far behind. The only competitor I've found is Salt, and I found its orchestration support to be a bit of a mess. And just as with Terraform, the code is constantly lagging behind the providers.

The one provider I'd have expected to be on the forefront of orchestration is Google, and in a different multiverse their engineers are swarming around Terraform to make sure it has top-notch, official, first-class support, but alas, not in this one.

Are there any competitors that provide a smoother experience?

Hey there- Dana from Google here, I lead the efforts around Terraform from our side. You'll be happy to know that in the last 2 months alone we've added import for:

  - google_bigquery_dataset
  - google_bigquery_table
  - google_compute_address
  - google_compute_disk
  - google_compute_global_address
  - google_compute_route
  - google_compute_network
  - google_dns_managed_zone
  - google_sql_user
  - google_storage_bucket
, with more to come shortly!

We only have one open issue around import, so if there are other resources you'd like to see imported feel free to file an issue: https://github.com/terraform-providers/terraform-provider-go... (just moved to a new repo a few days ago, and we're still in the process of getting existing issues moved over). A big factor in our prioritization of what to work on is based around issues filed (and thumbs ups on those issues), so that's a great way to get in touch with the team.

If you have any other questions around the Terraform+GCP experience, feel free to ask us in the #terraform channel in the GCP slack (https://gcp-slack.appspot.com/ if you aren't already there). Best of luck, and do reach out if you need anything!

Thanks! I'll take yet another look at Terraform, then. Fourth time the charm, or something.

My lasting fear is that even if it has 90% of the support, there will always be one thing, or one edge case or bug, that will become an annoying blocker. Using a tool like TF means becoming dependent on it to a large extent.

I've been using terraform for a couple of months now (love it), but honestly our biggest pain was just project organization. It looks like a lot of people make a file per-resource type (elb.tf, ec2.tf, rds.tf) but we thought that would be a lot of bloat. We opted to have a file per system (dev_db.tf, dev_ecs_asg.tf, dev_haproxy.tf, etc) and everything related to that particular system is contained in a single file (security groups, dns entries, roles/profiles, etc). But it's still in one flat directory per environment. (I know tf has introduced environments, but we haven't switched over yet.)

I know you can hack this together with modules, but it seems like environment/project organization would be easier if terraform just recursed subdirectories. Right? I've seen a couple of issues for it, but I don't believe I've seen a concrete reason why it's a no-go.

There's another similar issue with how EC2 security group rules are encoded: you can encode them either as ingress/egress stanzas on an "aws_security_group" resource, or you can attach rules to a security group resource with separate "aws_security_group_rule". You can't mix the two approaches on a single security group resource.

We adopted the ingress/egress stanza on security group resource approach.

If we ever wanted to change to the other approach (as described in the article), I don't think I would do state surgery by hand or even use "terraform state mv". I would:

    1. change terraforming to generate .tf files and tfstates the way I want
    2. remove the security groups from my config and my state
    3. use terraforming to regenerate the .tf files and tfstate

Hey there. Disclaimer: one of the creators of Terraform

I wanted to apologize that this is super confusing. All the scenarios where this exists (there are many) are historical. We originally went with the "nested" approach and now prefer the "split" approach for good reasons shown to us by users. But we kept both for backwards compatibility reasons. We have no good mechanism to enforce a migration at the moment. There are a couple ways we can resolve this technically in the future. For now, we should probably make sure the docs are annotated in all situations of the limitations of nested vs. standalone. I'll mention this to the team!

Yeah, this is a universal theme with Terraform resources.

If you can choose between a "con todo" description that lumps together resources and subresources, versus describing every single little thing separately and then attaching them together, always use attachments.

Oh, geeze. This is about: terraform.io

It would probably say Terraforming gotchas if it was about the other thing.

Ha! That was my first thought too. Where is the information about terraforming Mars?

Yeah, a title change would be nice. I was anticipating something much more interesting.

And I wanna hoverboard.

I absolutely can't stand how destructive terraform is by nature. We have switched to Ansible, which has an excellent AWS module, and never looked back.

What do you mean by destructive by nature? A resource will only be destroyed if:

1) You remove it explicitly from your configuration; which should be the intended result - it's no longer a "part of your infrastructure"

2) You've change some attribute that requires your provider to replace the instance entirely - Trying to think of an example off hand... While not applicable anymore as AWS now supports applying/changing IAM roles to running instances now, had you done something like that in the past, that would've forced a resource replacement.

Why Ansible rather than Elastic Beanstalk or Cloud Formation?

It's vendor-agnostic? And CF is super slow and limited to a small number of resources (not sure if it's a 100 or a 1000).

Definitely 100. And every little thing is a resource.

http://docs.aws.amazon.com/AWSCloudFormation/latest/UserGuid... Am I a missing something? Nothing in these docs about a 100 resource limit

It used to be 100. Looks like it's 200 now.

That's 200 resources per CF template. That's a massive CF template :)

I am just now implementing Packer and Vagrant in our devops workflows. Terraform is next on the list.

So far, it leaves me rather anxious - Packer and Vagrant appear to offer the bare minimum of usable functionality, with any advanced scenario bumping into (sometimes intentional) walls.

For example, it takes me 15-20 minutes to transfer a 50 MB file to a Windows VM being created by Packer. The GitHub issue, filed nearly 2 years ago, is closed with a comment that this is by design: https://github.com/hashicorp/packer/issues/2648#issuecomment...

Yet there is a PowerShell command that uses the same communication mechanism that can somehow do it in a matter of seconds. Of course, I cannot use this PowerShell command because Packer does not give me a variable with a machine's IP address because... it is improper somehow? https://github.com/hashicorp/packer/issues/4993

What the hell, Hashicorp...

I have a list of 10+ issues I have found so far and I am only starting to use these tools. From the activity in GitHub, they seem to be abandonware.

Maybe if I submitted PRs they might be accepted (then again, maybe not: https://github.com/hashicorp/packer/pulls) but I expect more from software than just accepting PRs - I expect its authors to actually develop it and to show an interest in improving it.

There is unfortunately nothing better out there. I admit, I am forced to use these products even though I do not find them satisfactory and the authors do not seem helpful.

If I had to start all over again with my current knowledge, I might perhaps just write my own scripting and skip Packer/Vagrant altogether. The value they offer with VM management comes with the downside of being left in the mud and having the system work against you when you try something nontrivial.

I am scared of what I will find when I touch Terraform. As I write this, I think I will first see whether I can just script it manually.

Hey author here! Happy to answer any questions etc :-)

No questions, just a suggestion: implement the part where the terraform plan is added as a comment in the PR. We set this up at my current employer and it makes the review process much quicker (also, commenting on lines in the ~plan~ terraform code changes is the bee's knees).

Don't have the apply be automatic after a review is approved; terraform apply's occasionally go sideways and need human intervention (remember: rollbacks are not automatic). A human should always kick off the apply and monitor state change activity.

Woah can you comment on lines in a comment? /me dashes to investigate github

And thanks for the suggestion. So far it's been on a someday maybe list, but if it really does help that much, maybe we'll bump it to someday maybe soon.

Questions from our team:

- Are you commenting with the output of show on the planfile to get human-readable version?

- Line by line commenting on comments?

- Do you have state-splits? Do you run plan on each individually for every PR?

May I email you answers to these?

Yes, that works. (email in profile)

I couldn't agree more with this, we do exactly the same and it works great for us. Plan on a comment and manual apply.

Great post. I think the area where things start to go off the rails with Terraform is variables/modules though. There are edge cases and gotchas with modules and not a lot of best practice guides.

Shameless plug, if anybody is looking for a guide on getting started with Terraform and Google Cloud Platform, I wrote a short blog post:


Shameless plug, if anybody wants to generate a Terraform module with a testing harness preconfigured then I hacked up a module scaffolder last night/this morning: https://github.com/howdoicomputer/terrashovel

Great post! I too had the issue with AWS volume attachments and the need to separate them out. Luckily it happened early during some of my first provisioning and we realized it before standing up more databases. I now have a policy where each server has a separate EBS volume, using a volume attachment Terraform resource, for any data. Inline EBS volumes are just used for the OS.

I've run into another similar type issue now and I think that I'm going to have to do state surgery like you. I want to refactor some of my terraform resources into modules, but this changes the resource names. I can't see any way to do it without either standing up new infrastructure or modifying the state directly.

We ran into similar issues. From there I went one step further and wrote a tool to do the attachments for me (instead of using Terraform's aws_volume_attachment):


This way in Terraform I provision a group of volumes, a group of instances, and rely on `goat` to do the rest.

That's pretty cool. So is the model that you don't manage the volumes in TF at all?

(Btw I think I know people you know... also in Montreal and I know a couple folks at AdGear!)

That's cool, I'll ask my coworkers.

And yes this way I keep EBS and the rest of my recipes as far away as possible.

Something like:

``` kafka-volumes/ <ebs volumes> terraform.tfstate kafka/ (use local backend ../kafka-volumes/terraform.tfstate) <ec2 instances> ```

We're using Terraform for our AWS env and while it worked great for getting stuff out there, we're now scrambling to figure out how to get Terraform to do non-destructive updates to the environment. For instance, I need to update our base OS for the application servers we're running. How do I do this without incurring a downtime? Or I need to now interject a new reverse proxy between our ELB's and app servers, but want to do it without causing an interruption. We're doing our research and this blog post is very helpful, but if you have any pointers...

You pretty much have to take the same approach as you would outside of terraform: create the new thing (launch configuration, standalone instance) and attach it to the ELB before spinning down the old one. Terraform does have some rudimentary aid to that end in the form of the create_before_destroy[0] flag, though it doesn't work out with uniquely named things.

[0] https://www.terraform.io/docs/configuration/resources.html#c...

You should use 2 ASG ( blue / green ) and 1 ELB for that, then when you need to update the OS of one of the ASG you just update the launch configuration. It's easy to do in Terraform.

1) Update the launch configuration of your un-used ASG ( with the new AMI ID )

2) Apply terraform to deploy the new ASG

3) Make sure it's working ( your local app on your new instances )

4) Connect the ELB to your new ASG

5) Set the old ASG to 0 instances to drain the connections

This is a pretty common question. Check out this blog post for how to get Terraform to bring up a new ASG with a new AMI before removing the old one from an ELB:


hi Kalmar,

Is it possible to query information directly from a state file? Some interpolation would help. I ask this because I often refer to resource information from another project, and I don't see that modules are helpful.

I posted my work-around here [1] some days ago . The real code was not there because I haven't had permission for my company, but it's very short and easy to write by anyone.

Thanks for your reading.

[1] https://github.com/icy/oops/blob/master/terraform/problem_1_...

Have you tried Terragrunt[1]?

[1] https://github.com/gruntwork-io/terragrunt

All the useful things terragrunt does have been built in since Terraform 0.9.

Disclaimer: ex-HashiCorp employee.

Not true.

Terraform v0.9 only handles remote state and locking properly. Handling multiple applies of state is something scheduled for 0.10

Basically terragrunt handles some stuff today in a less robust, less portable, way that your team is likely to need today -- whilst waiting for core terraform support.

(disclaimer: I follow terragrunt and terraform very closely but do not actually use terragrunt in producution anymore).

Your site appears to be down, the connection is timing out for me

here's something i got bit by more recently re: terraform plan -out and tooling using Terraform's Golang API.

Handling package dependencies with Go is not straighforward. There are several ways of doing it, and none are native to Golang.

Additionally, Go doesn't support getting versions of packages by tag or branch.

This bit me hard when I tried to update Palantir's TFJSON utility (turns tfplan binaries into json) so I could do unit testing of my Terraform plans with rspec.

The utility depended on v0.7.4 of terraform, but Terraform maintains a plan format constant that defines which plans can be used by what versions. They changed the plan format between 0.7.4 and 0.9.8 without bumping that constant, so when I tried running tfjson against plans created by the latter version, I got a weird non-matching datatype error that took a while to figure out. (I eventually had to vimdiff the hex outputs of plans created by both versions to figure that out.)

Additionally, HashiCorp made a significant change to the way they handled providers between 0.9.8 and 0.10.0 that justified them to bump the plab format version AGAIN. The catch: 0.10.0 isn't released yet, despite that being the code in their master branch.

I figured that updating tfjson's vendored terraform library to 0.9.8 would solve it. I first did a go get to fetch the latest TF codebase and used gvt to vendor it. That's when I discovered that plans generated by 0.9.8 are no longer compatible. After discovering that go get can't fetch packages by tag (Hashicorp tags their release commita) because Google believes in stable HEADs, I had to find a tool that could support fetching packages by tags. Govendor did that, so I used that.

It takes FOREVER to fetch all of the subpackages used by terraform. I couldn't do it during a three hour flight. Rubygems has its problems, but fetching deps isn't one of them. And even when I thought I fetched the entire source tree at v0.9.8, I would still get errors about missing types or missing packages.

I'm hopeful that I'll eventually find a solution, but it's a dog compared to using Gemfile.lock.

I thought terragrunt was a must have for that kind of deployment.

I'm curious about that as well. I was told by coworkers that Hashicorp added support for DynamoDB which rendered terragrunt redundant but I haven't had time to look into it.

terraform {

  backend "s3" {

    region = "us-west-1"

    bucket = "foo-tf-us-west-1"

    key = "foobar.tfstate"

    dynamodb_table = "tf-lock"



Awesome! Any idea when this was added? I feel like this wasn't in the documentation a week or two ago; everything had still said "Use Hashicorp Atlas for remote state locking".

I've been using it for weeks if not months.

Their documentation might be lagging occasionally. It's a small team tackling a big challenge.

This sums up our experience with Terraform perfectly:

> Most outages are caused by human error and configuration changes, and applying Terraform changes is a terrifying mix of the two.

Terraform is a great tool nonetheless. Just like Heap, we have code reviews for the configuration itself, and a CI pipeline for validating it. This pipeline is quite superficial (`terraform validate` mostly does syntax checking), so we are too working on using centralized state to `terraform plan` for reviews.

>Terraform state surgery

Did you try to use `terraform state mv`? I've found that command useful (albeit for much less than thousands of resources).

I don't think it would work in this case, as the `ebs_block_device` block isn't a resource. In fact, the TF state doesn't even have the volume IDs for them!

An alternative to doing this was `terraform import` on all the volumes, then defining attachments, and hoping it all worked when you run `terraform plan`. I don't 100% remember now why we didn't do that.

So... ahem what other idiots came here expecting a post about troubleshooting Martian habitability?

::Sulks off dejectedly::

Wikipedia has a decent article if you've never read it: https://en.wikipedia.org/wiki/Terraforming_of_Mars#Challenge...

Now that project would be cool to see.

At least one :)

at least two :)

at least many!

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact