Add available memory state for hostmetrics receiver

[rogercoll]

The current Linux memory states provided by the hostmetricsreceiver might be not the most appropriate ones for an end user. The available memory states for Linux are:

[buffered, cached, free, slab_reclaimable, slab_unreclaimable, used]

All of those values are retrieved with gopsutil, which gets them from /proc/meminfo file.

A user might think that the free value represents the amount of memory that the system can use, but free (MemFree) is a count of how many pages are free in the buddy allocator, however there are a lot of ways pages can be returned to the buddy allocator in time of need. The latest is provided by the Linux kernel (>3.16) as MemAvailable in /proc/meminfo.

Can we get MemAvailable field from current states?

No, it is an independent field not computed with the current values. Gopsutils provides this value regardless the kernel version as it can be estimated from MemFree, Active(file), Inactive(file), and SReclaimable, as well as the "low" watermarks from /proc/zoneinfo.

Repercussion

The user will be able to know the actual available memory of the system.

• Will it break the assumption that the sum of all metrics states should be equal to the limit (total memory)? https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/metrics/semantic_conventions/README.md#instrument-naming

It is already broken and quite difficult to achieve it as some of the memory state values are inclusive, for example the cached state. The assumption would be defined as:

total = buffered + cached + free + used + slab_reclaimable + slab_unreclaimable

But the used value is computed by gopsutils as: total - free - buffered - cached

Then:

total = buffered + cached + free + (total - free - buffered - cached) + slab_reclaimable + slab_unreclaimable

total = total + slab_reclaimable + slab_unreclaimable

The sum(state) will always be higher than the total.

(The following proposals do not solve the sum assumption, they only add more memory information for the user)

Proposal 1

Add a new memory state for Linux to provide the available memory. State name: avaialble

Proposal 2

Change used state for: total - available
This proposal requires to add the memory.limit metric as we won't be able to compute the total memory from the states. Currently, we can get the total amount of memory with the following promQL query: scalar(sum without (state) (system_memory_usage{state =~ "free|used|cached|buffered"}))

[jpkrohling]

cc @dmitryax

[bogdandrutu]

/cc @dashpole I think you are using this as well

[quentinmit]

I don't think either of those proposals improves the problem with the values not adding up. I'd like to add proposal 3, fixing the set of states to exactly cover the memory on the system. If there are states that double-count memory, they should be separate metrics. Here are the non-overlapping states from Munin and how they calculate them from /proc/meminfo as an inspiration:

apps (MemTotal-MemFree-Buffers-Cached-Slab-PageTables-Percpu-SwapCached)
buffers (Buffers)
swap (SwapTotal-SwapFree)
cached (Cached-Shmem)
free (MemFree)
shmem (Shmem)
slab_cache (Slab)
swap_cache (SwapCached)
page_tables (PageTables)
per_cpu (Percpu)
vmalloc_used (VmallocUsed)

(I'm not sure if it makes sense to include swap; I'm guessing that with the current set of states we're not including swap? You could leave that out and I think the remaining states would represent total physical memory.)

I think the main difference between this and proposal 2 is that here we're changing the definition of used (what Munin calls apps) instead of changing the definition of free.

[rogercoll]

@quentinmit I really like your proposal to solve the sum assumption, it provides much more information to the user and we prevent having overlapping states.

Tested without the Swap states as they are covered in the paging scraper and the sum usage is correct. Nonetheless, I have a couple of concerns:

• The scraper relies on gopsutils for those metrics and the library is not currently providing percpu state (I can do the PR)
• What about the "available" memory state that the kernel provides (MemAvailable)? (available != total - apps)

In my opinion, the available or non-available state is very useful metric when defining alerts. The metric used (or apps) should not be understood as non-available memory, usually, the latest is quite higher.

[rogercoll]

Proposal 3

Another possible solution would be to add a new memory metric only for this purpose, as making any change on the current states would break the sum assumption.

The memory system scrapper will generate the following metrics:

- system.memory.usage: state attributes
- system.memory.utilization: disabled by default, state attributes
- system.memory.available: no state attributes

I would appreciate any opinion on this

[rogercoll]

Whether using the used state memory or the total - available memory can be truly significant in alerting systems. I made a small test to check the average variation between both values while stressing the memory:

The average difference is around 7/8% . By providing both values we let the user choose whatever it's best.

[TylerHelmuth]

Another possible solution would be to add a new memory metric only for this purpose

Is this the same as your proposed solution 1?

I like the idea of adding a new metric; allows those who currently rely on used to continue to use it, while providing a more accurate measure for those who want it.

[rogercoll]

@TylerHelmuth No, thanks for pointing that out. It would be a new proposal, let me update the comment.

Proposals 1 and 2 are based on the "state" attribute used in usage and utilization metrics. Proposal 3 won't do any modification on the state neither the current metrics, instead, add a new one.

[dmitryax]

Adding another metric system.memory.available maybe be a good solution to one the problems described in this issue. But it's still unclear how to solve the total mismatch issue and also adding this metric brings confusion to end user questioning what's the difference between system.memory.available and system.memory.usage.

Potentially, we can solve both problems by introducing a concept of optional metric attribute values to the metrics builder. The attributes values can be enabled or disabled in metadata.yaml (default) and user configs. If attribute value is disabled, it means that datapoints with this attribute value won't be emitted.

With this functionality added to the metrics builder, we will have the following sets of state attribute values in this receivers:

• Enabled by default: buffered, cached, free and used. Sums up to a correct total value.
• Disabled by default: slab_reclaimable, slab_unreclaimable and available. Optional values that user can enable.

What do you think?

@bogdandrutu @djaglowski please share your opinion on the additional feature to the metrics builder

[djaglowski]

I think we should aim for a design that adheres to the specification's aggregation expectations (ie. data points sum to a meaningful total). Supporting metrics that violate these expectations seems like the wrong approach to me.

There seem to be two layers to the problem here:

1. What is the correct data model for this situation?
2. How do we migrate from a flawed data model to the correct one?

I don't have a strong opinion on the first question, but in regards to the second, I think we basically have to decide whether to fix it, or to deprecate and replace it. The right approach depends on stability level. In my opinion, a per-metric stability level is necessary in order to have a clear approach (eg. alpha -> beta -> stable -> deprecated). If this metric is less than stable, then we should fix it. If it is "stable" but realized to be fundamentally flawed in its design, we should deprecate it and add one or more new metrics that amount to a more accurate model.

[dmitryax]

I think we should aim for a design that adheres to the specification's aggregation expectations (ie. data points sum to a meaningful total). Supporting metrics that violate these expectations seems like the wrong approach to me.

That's true, we should always aim to have meaningful total values. Proposing optional attribute values, I imply that the sum of attribute values enabled by default should represent the most reasonable total value. At the same time, I can think of use cases when user wants to disable some attributes values or add others optional ones and have their own custom meaningful sums out of them. In this case, moving slab_reclaimable and slab_unreclaimable to a different metric to get correct total value would significantly reduce flexibility of using them because they are part of the same memory usage metric.