Architecture

This section is intended to give you an overview of the architecture of Krill, which is important to keep in mind when deploying the application in your infrastructure. It will give you an understanding how and where data is stored, how to make your setup redundant and how to save and restore backups.

Warning

Krill does NOT support clustering at this time. You can achieve high availability by doing a fail-over to a standby inactive installation using the same data and configuration. However, you cannot have multiple active instances.

Used Disk Space

Krill stores all of its data under the DATA_DIR. For users who will operate a CA under an RIR / NIR parent the following sub-directories are relevant:

Directory	Contents
data_dir/ssl	The HTTPS key and certificate used by Krill
data_dir/cas	The history of your CA(s) in raw JSON format
data_dir/pubd	If used, the history of your Publication Server

Note

Note that old versions of Krill also used the directories data_dir/rfc8181 and data_dir/rfc6492 for storing all protocol messages exchanged between your CAs and their parent and repository. If they are still present on your system, you can safely remove them and save space - potentially quite a bit of space.

Archiving

Krill offers the option to archive old, less relevant, historical information related to publication. You can enable this by setting the option archive_threshold_days in your configuration file. If set Krill will move all publication events older than the specified number of days to a subdirectory called archived under the relevant data directory, i.e. data_dir/pubd/0/archived if you are using the Krill Publication Server and data_dir/cas/<your-ca-name>/archived for each of your CAs.

You can set up a cron job to delete these events once and for all, but we recommend that you save them in long term storage if you can. The reason is that if (and only if) you have this data, you will be able to rebuild the complete Krill state based on its audit log of events, and irrevocably prove that no changes were made to Krill other than the changes recorded in the audit trail. We have no tooling for this yet, but we have an issue on our backlog.

Saving State Changes

You can skip this section if you’re not interested in the gory details. However, understanding this section will help to explain how backup and restore works in Krill, and how a standby fail-over node can be used.

State changes in Krill are tracked using events. Krill CA(s) and Publication Servers are versioned. They can only be changed by applying an event for a specific version. An event just contains the data that needs to be changed. Crucially, they cannot cause any side effects. As such, the overall state can always be reconstituted by applying all past events. This concept is called event-sourcing, and in this context the CAs and Publication Servers are so-called aggregates.

Events are not applied directly. Rather, users of Krill and background jobs will send their intent to make a change through the API, which then translates this into a so-called command. Krill will then lock the target aggregate and send the command to it.

Upon receiving a command the aggregate (your CA etc.) will do some work. In some cases a command can have a side-effect. For example it may instruct your CA to create a new key pair, after receiving entitlements from its parent. The key pair is random — applying a command again would result in a new random key pair. Remember that commands are not re-applied to aggregates, only their resulting events are. Thus in this example there would be an event caused that contains the resulting key pair.

After receiving the command, the aggregate will return one of the following:

1. An error

   Usually this means that the command is not applicable to the aggregate state. For example, you may have tried to remove a ROA which does not exist.

   When Krill encounters such an error, it will store the command with some meta-information like the time the command was issued, and a summary of the error, so that it can be seen in the history. It will then unlock the aggregate, so that the next command can be sent to it.

2. No error, zero events

   In this case the command turned out to be a no-op, and Krill just unlocks the aggregate. The command sequence counter is not updated, and the command is not saved. This is used as a feature whenever the ‘republish’ background job kicks in. A ‘republish’ command is sent, but it will only have an actual effect if there was a need to republish — e.g. a manifest would need to be re-issued before it would expire.

3. One or more events

   In this case there is a desired state change in a Krill aggregate. Krill will now apply and persist the changes in the following order:

   • Each event is stored. If an event already exists for a version, then then the update is aborted. Because Krill cannot run as a cluster, and it uses locking to ensure that updates are done in sequence, this will only fail on the first event if a user tried to issue concurrent updates to the same CA.

   • On every fifth event a snapshot of the state is saved to a new file. If this is successful then the old snapshot (if there is one) is renamed and kept as a backup snapshot. The new snapshot is then renamed to the ‘current’ snapshot.

   • When all events are saved, the command is saved enumerating all resulting events, and including meta-information such as the time that the time that the command was executed. And when multiple users will be supported, this will also include who made a change.

   • Finally the version information file for the aggregate is updated to indicate its current version, and command sequence counter.

Warning

Krill will crash, by design, if there is any failure in saving any of the above files to disk. If Krill cannot persist its state it should not try to carry on. It could lead to disjoints between in-memory and on-disk state that are impossible to fix. Therefore, crashing and forcing an operator to look at the system is the only sensible thing Krill can now do. Fortunately, this should not happen unless there is a serious system failure.

Loading State at Startup

Krill will rebuild its internal state whenever it starts. If it finds that there are surplus events or commands compared to the latest information state for any of the aggregates, then it will assume that they are present because, either Krill stopped in the middle of writing a transaction of changes to disk, or your backup was taken in the middle of a transaction. Such surplus files are backed up to a subdirectory called surplus under the relevant data directory, i.e. data_dir/pubd/0/surplus if you are using the Krill Publication Server and data_dir/cas/<your-ca-name>/surplus for each of your CAs.

Recover State at Startup

When Krill starts, it will try to go back to the last possible recoverable state if:

• it cannot rebuild its state at startup due to data corruption

• the environment variable: KRILL_FORCE_RECOVER is set

• the configuration file contains always_recover_data = true

Under normal circumstances, i.e. when there is no data corruption, performing this recovery will not be necessary. It can also take significant time due to all the checks performed. So, we do not recommend forcing this.

Krill will try the following checks and recovery attempts:

• Verify each recorded command and its effects (events) in their historical order.

• If any command or event file is corrupt it will be moved to a subdirectory called corrupt under the relevant data directory, and all subsequent commands and events will be moved to a subdirectory called surplus under the relevant data directory.

• Verify that each snapshot file can be parsed. If it can’t then this file is moved to the relevant corrupt sub-directory.

• If a snapshot file could not be parsed, try to parse the backup snapshot. If this file can’t be parsed, move it to the relevant corrupt sub-directory.

• Try to rebuild the state to the last recoverable state, i.e. the last known good event. Note that if this pre-dates the available snapshots, or, if no snapshots are available this means that Krill will try to rebuild state by replaying all events. If you had enabled archiving of events, it will not be able rebuild state.

• If rebuilding state failed, Krill will now exit with an error.

Note that in case of data corruption Krill may be able to fall back to an earlier recoverable state, but this state may be far in the past. You should always verify your ROAs and/or delegations to child CAs in such cases.

Of course, it’s best to avoid data corruption in the first place. Please monitor available disk space, and make regular backups.

Backup / Restore

Backing up Krill is as simple as backing up its data directory. There is no need to stop Krill during the backup. To restore put back your data directory and make sure that you refer to it in the configuration file that you use for your Krill instance. As described above, if Krill finds that the backup contain an incomplete transaction, it will just fall back to the state prior to it.

Warning

You may want to encrypt your backup, because the data_dir/ssl directory contains your private keys in clear text. Encrypting your backup will help protect these, but of course also implies that you can only restore if you have the ability to decrypt.

Krill Upgrades

All Krill versions 0.4.1 and upwards can be automatically upgraded to the current version. Any required data migrations will be performed automatically. To do so we recommend that you:

• backup your krill data directories

• install the new version of Krill

• stop the running Krill instance

• start Krill again, using the new binary, and the same configuration

If you want to test if data migrations will work correctly for your data, you can do the following:

• copy your data directory to another system

• set the env variable KRILL_UPGRADE_ONLY=1

• create a configuration file, and set data_dir=/path/to/your/copy

• start up Krill

Krill will then perform the data migrations, rebuild its state, and then exit before doing anything else.

Krill Downgrades

Downgrading Krill data is not supported. So, downgrading can only be achieved by installing a previous version of Krill and restoring a backup from before your upgrade.

Proxy and HTTPS

HTTPS Mode

Krill uses HTTPS by default, and will generate a key pair and create a self-signed certificate if no previous key pair or certificate is found. Files are stored under the data directory as ssl/key.pem and ssl/cert.pem respectively.

Alternatively you make Krill configure krill to not generate these files but use existing files at the same file locations. This should work, but has not been tested extensively. To use this mode you can use `https_mode = "existing"` in your krill configuration file.

It also possible to force Krill to disable HTTPS and use plain HTTP. We do not recommend this set up, but it may be useful in certain setups. Arguably, as long as Krill listens on 127.0.0.1 only (as is the default), and an HTTPS enabled proxy server is used for public access, then having plain HTTP traffic between the proxy and Krill over the loopback interface is not necessarily problematic. To use this mode set `https_mode = "disable"` in your configuration file.

If you need to access the Krill UI or API (also used by the CLI) from another machine, then we highly recommend that you use a proxy server such as NGINX or Apache. This proxy can then also use a proper HTTPS certificate signed by a web TA, and production grade TLS support.

Proxy Krill UI

The Krill UI and assets are hosted directly under the base path /. So, in order to proxy to the Krill UI you should proxy ALL requests under / to the Krill back-end.

Note that although the UI and API are protected by a token, you should consider further restrictions in your proxy setup, such as restrictions on source IP or adding your own authentication.

Proxy Krill as Parent

If you delegated resources to child CAs then you will need to ensure that these children can reach your Krill. Child requests for resource certificates are directed to the /rfc6492 directory under the service_uri that you defined in your configuration file.

Note that contrary to the UI you should not add any additional authentication mechanisms to this location. RFC 6492 uses cryptographically signed messages sent over HTTP and is secure. However, verifying messages and signing responses can be computationally heavy, so if you know the source IP addresses of your child CAs, you may wish to restrict access based on this.

Proxy Krill as Publication Server

If you are running Krill as a Publication Server, then you should read here how to do the Publication Server specific set up.

Warning

We recommend that you do not make Krill available to the public internet unless you really need remote access to the UI or API, or you are serving as parent CA or Publication Server for other CAs.