fleet/handbook/engineering/scaling-fleet.md

Scaling Fleet

This is a document that evolves and will likely always be incomplete. If you feel like something is missing, either add it or bring it up in any way you consider.

What have we learned so far?

• How Fleet scales
• How to prevent most of this
• Foreign keys and locking
• Insert on duplicate update
• Host extra data and JOINs
• What DB tables matter more when thinking about performance?
• Expose more host data in the host listing
• Understand main use-cases for queries
• On constantly changing data
• Counts and aggregated data
• Caching data such as app config
• Redis SCAN
• Connecting to Dogfood MySQL & Redis

How Fleet scales

Nowadays, Fleet, as a Go server, scales horizontally very well. It’s not very CPU or memory intensive. In terms of load in infrastructure, from highest to lowest are: MySQL, Redis, and Fleet.

In general, we should burn a bit of CPU or memory on the Fleet side if it allows us to reduce the load on MySQL or Redis.

In many cases, caching helps, but given that we are not doing load balancing based on host id (i.e., make sure that the same host ends up in the same Fleet server). This goes only so far. Caching host-specific data is not done because round-robin LB means all Fleet instances end up circling the total list of hosts.

How to prevent most of this

The best way we’ve got so far to prevent any scaling issues is to load test things. Every new feature must have its corresponding osquery-perf implementation as part of the PR, and it should be tested at a reasonable scale for the feature.

Besides that, you should consider the answer(s) to the following question: how can I know that the feature I’m working on is working and performing well enough? Add any logs, metrics, or anything that will help us debug and understand what’s happening when things unavoidably go wrong or take longer than anticipated.

HOWEVER (and forgive this Captain Obvious comment): do NOT optimize before you KNOW you have to. Don’t hesitate to take an extra day on your feature/bug work to load test things properly.

Foreign keys and locking

Among the first things you learn in database data modeling is: that if one table references a row in another, that reference should be a foreign key. This provides a lot of assurances and makes coding basic things much simpler.

However, this database feature doesn’t come without a cost. The one to focus on here is locking, and here’s a great summary of the issue: https://www.percona.com/blog/2006/12/12/innodb-locking-and-foreign-keys/

The TLDR is: understand very well how a table will be used. If we do bulk inserts/updates, InnoDB might lock more than you anticipate and cause issues. This is not an argument to not do bulk inserts/updates, but to be very careful when you add a foreign key.

In particular, host_id is a foreign key we’ve been skipping in all the new additional host data tables, which is not something that comes for free, as with that, we have to keep the data consistent by hand with cleanups.

Insert on duplicate update

It’s very important to understand how a table will be used. If rows are inserted once and then updated many times, an easy reflex is to do an INSERT … ON DUPLICATE KEY UPDATE. While technically correct, it will be more performant to try to do an update, and if it fails because there are no rows, then do an insert for the row. This means that it’ll fail once, and then it’ll update without issues, while on the INSERT … ON DUPLICATE KEY UPDATE, it will try to insert, and 99% of the time, it will go into the ON DUPLICATE KEY UPDATE.

This approach has a caveat. It introduces a race condition between the UPDATE and the INSERT where another INSERT might happen in between the two, making the second INSERT fail. With the right constraints (and depending on the details of the problem), this is not a big problem. Alternatively, the INSERT could be one with an ON DUPLICATE KEY UPDATE at the end to recover from this scenario.

This is subtle, but an insert will update indexes, check constraints, etc. At the same time, an update might sometimes not do any of that, depending on what is being updated.

While not a performance GOTCHA, if you do use INSERT … ON DUPLICATE KEY UPDATE, beware that LastInsertId will return non-zero only if the INSERT portion happens. If an update happens, the LastInsertId will be 0.

Host extra data and JOINs

Indexes are great. But like most good things, the trick is in the dosage. Too many indexes can be a performance killer on inserts/updates. Not enough, and it kills the performance of selects.

Data calculated on the fly cannot be indexed unless it’s precalculated (see counts section below for more information).

Host data is among the data that changes and grows the most in terms of what we store. It used to be that we used to add more columns in the host table for the extra data in some cases.

Nowadays, we don’t update the host table structure unless we really, really, REALLY need to. Instead, we create adjacent tables that reference a host by id (without an FK). These tables can then be JOINed with the host table whenever needed.

This approach works well for most cases, and for now, it should be the default when gathering more data from hosts in Fleet. However, it’s not a perfect approach as it has its limits.

JOINing too many tables, sorting based on the JOINed table, etc., can have a big performance impact on selects.

Sometimes one strategy that works is selecting and filtering the adjacent table with the right indexes; then, JOIN the host table to that. This works when only filtering/sorting by one adjacent table and pagination can be tricky.

Solutions can become a curse too. Be mindful of when we might cross that threshold between good and bad performance.

What DB tables matter more when thinking about performance?

While we need to be careful about handling everything in the database, not every table is the same. The host and host_* tables are the main cases where we have to be careful when using them in any way.

However, beware of tables that go through async aggregation processes (such as scheduled_query and scheduled_query_stats) or those that are read often as part of the osquery distributed/read and config endpoints.

Expose more host data in the host listing

Particularly with extra host data (think MDM, Munki, Chrome profiles, etc.), another GOTCHA is that some users have built scripts that go through all hosts by using our list host endpoint. This means that any extra data we add might cause this process to be too slow or timeout (this has happened in the past).

Beware of this, and consider gating the extra data behind a query parameter to allow for a performant backward compatible API that can expose all the data needed in other use cases.

Calculated data is also tricky in the host listing API at scale, as those calculations have to happen for each host row. This can be extra problematic if the sort happens on the calculated data, as all data has to be calculated across all hosts before being able to sort and limit the results (more on this below).

Understand main use-cases for queries

Be aware of the use cases for an API. For example, take the software listing endpoint. This endpoint lists software alongside the number of hosts with that item installed. It was designed to be performant in a limited use case: list the first eight software items, then count hosts for those software ids.

The problem came later when we learned that we missed an important detail: the UI wanted to sort by amount of host count so that the most popular software appeared on top of this.

This resulted in basically a full host_software table scan per each software row to calculate the count per software. Then, sort and limit. The API worked in the simple case, but it timed out for most customers in the real world.

On constantly changing data

It can be difficult to show real-time presence data. For Fleet, that is the host seen_time -- the time a host last connected to the server -- which is used to determine whether a host is online.

Host seen_time is updated with basically every check-in from any kind of host. Hosts check in every 10 seconds by default. Given that it’s a timestamp reflecting the last time a host contacted Fleet for anything, it’s always different.

While we are doing a few things to make this better, this is still a big performance pain point we have. In particular, we are updating it in bulk. It used to be a column of the hosts' table, which caused a lot of locking. Now it’s an adjacent table without FK.

Luckily, we don’t have anything else (at least up to the moment of this writing) that changes as often as seen_time. However, other features such as software last used can cause similar issues.

Counts and aggregated data

UX is key for any software. APIs that take longer than 500ms to respond can cause UX issues. Counting things in the database is a tricky thing to do.

In the ideal case, the count query will be covered by an index and be extremely fast. In the worst case, the query will be counting filtering on calculated data, which results in a full (multi) table scan on big tables.

An approach we've taken to addressing this is pre-calculating aggregations and counts that take a long time to generate. By generating these results beforehand and storing them, we can return results by reading a single row from a table when the information is needed.

This approach has a handful of issues:

• The accuracy of the data is worse. We will never get truly accurate counts (the “real-time” count the API returns could change 1ms after we get the value).
• Depending on how many ways we want to count things, we will have to calculate and store them.
• Communicating to the user the interval at which things update can sometimes be tricky.

All of this said, Fleet and osquery work in an “update at an interval” fashion, so we have exactly one pattern to communicate to the user, and then they can understand how most things work in the system.

Caching data such as app config

Caching is a usual strategy to solve some performance issues in the case of Fleet level data, such as app config (of which we will only have one), is easy, and we cache at the Fleet server instance level, refreshing the value every one second. App config gets queried with virtually every request, and with this, we reduce drastically how many times the database is hit with that query. The side effect is that a configuration would take one second to be updated in each Fleet instance, which is a price we are willing to pay.

Caching host-level data is a different matter, though. Given that Fleet is usually deployed in infrastructure where the load balancer distributes the load in a round-robin-like fashion (or maybe other algorithms, but nothing aware of anything within Fleet itself). Then virtually all hosts end up being seen by all Fleet instances, so caching host-level data (in the worst case) results in having a copy of all the hosts in each Fleet instance and refreshing that at an interval.

Caching at the Fleet instance level is a great strategy if it can reasonably handle big deployments, as Fleet utilizes minimal RAM.

Another place to cache things would be Redis. The improvement here is that all instances will see the same cache. However, Redis can also be a performance bottleneck depending on how it’s used.

Redis SCAN

Redis has solved many scaling problems in general, but it’s not devoid of scaling problems of its own. In particular, we learned that the SCAN command scans the whole key space before it does the filtering. This can be very slow, depending on the state of the system. If Redis is slow, a lot suffers from it.

Connecting to Dogfood MySQL & Redis

When investigating performance issues, it can be helpful to connect directly to the MySQL and Redis instances to run queries and inspect data. Below are instructions for connecting to the Dogfood MySQL and Redis instances.

Prerequisites

1. Setup VPN
2. Configure SSO

MySQL

Get the database host:

DB_HOST=$(aws rds describe-db-clusters --filter Name=db-cluster-id,Values=fleet-dogfood --query "DBClusters[0].Endpoint" --output=text)

Get the database user:

DB_USER=$(aws rds describe-db-clusters --filter Name=db-cluster-id,Values=fleet-dogfood --query "DBClusters[0].MasterUsername" --output=text)

Get the database password:

DB_PASSWORD=$(aws secretsmanager get-secret-value --secret-id fleet-dogfood-database-password --query "SecretString" --output=text)

Connect:

mysql -h"${DB_HOST}" -u"${DB_USER}" -p"${DB_PASSWORD}"

Redis

Get the Redis Host:

REDIS_HOST=$(aws elasticache describe-replication-groups --replication-group-id fleetdm-redis --query "ReplicationGroups[0].NodeGroups[0].PrimaryEndpoint.Address" --output=text)

Connect:

redis-cli -h "${REDIS_HOST}"