inputs.zfs: add Linux txgs metrics collection

[mehulanshumali]

Use Case

On Linux, ZFS exposes transaction group history in /proc/spl/kstat/zfs/<pool>/txgs.

This file contains useful pool-level write activity and timing information, including:
txg, birth, ndirty, nread, nwritten, reads, writes, otime, qtime, wtime, and stime.

The current inputs.zfs plugin does not collect these metrics, so Telegraf users cannot observe txg behavior directly even though the data is available from the kernel.

This would help monitor ZFS write pressure, txg progression, and timing behavior during real workloads.

Expected behavior

Add optional Linux-only support in inputs.zfs to collect txg metrics from /proc/spl/kstat/zfs/<pool>/txgs.

Suggested behavior:

• feature disabled by default
• collect only the latest completed txg row per pool
• emit a zfs_txg measurement
• add pool as a tag
• expose numeric txg columns as fields
• allow optional pool selection through config

Example:

[[inputs.zfs]]
  txgMetrics = true
  txgPools = ["pool1", "pool2"]

Actual behavior

inputs.zfs currently does not read txgs files, so these metrics are unavailable in Telegraf.

Additional info

No response

[mehulanshumali]

I have a concrete implementation plan for this feature and I’d like to work on it.

Proposed flow:

• add Linux-only support in inputs.zfs
• keep it disabled by default via txgMetrics = false
• support optional txgPools filtering:
  • if txgPools is set, only read the listed pool paths, for example:
    • txgPools = ["tank", "backup"]
    • Telegraf would read:
      • /proc/spl/kstat/zfs/tank/txgs
      • /proc/spl/kstat/zfs/backup/txgs
  • if txgPools is not set, auto-discover pool txg files using the wildcard path /proc/spl/kstat/zfs/*/txgs, where * matches the pool name
• for each txgs file, parse the header and select only the latest completed row (state == "C")
• suppress duplicate emission across scrapes by tracking the last emitted completed txg row per pool and only emitting when the selected row changes
• emit one zfs_txg metric per pool with tag pool and the numeric txg columns as fields

This keeps the collection focused on completed txgs, ignores older history rows in the file, and avoids re-emitting the same completed txg on every scrape.

If this looks good, I can work on the implementation and open a PR.

[srebhan]

Thanks for bringing this up @mehulanshumali!

• keep it disabled by default via txgMetrics = false
• support optional txgPools filtering:

I would suggest to either

• add a setting includes with a setting transactions enabling the collection of transaction statistics for all pools and rely on metric filtering

or

• add the transaction filter list (e.g. transaction_pools) and disable transaction collection if the list is empty

This avoids the redundancy and the corner-case where the filter has entries but the global setting is false.

• for each txgs file, parse the header and select only the latest completed row (state == "C")
• suppress duplicate emission across scrapes by tracking the last emitted completed txg row per pool and only emitting when the selected row changes

Hmmm, from what I understand you are trying to avoid emitting already emitted transaction groups, don't you?
If so, your approach will result in an unexpected sampling of the transaction groups as you seem to emit only the latest transaction group. If people want to see all groups the would need to set the interval of the plugin to insanely short values.
How about keeping track of the last (unique, strictly monotonic) transaction-group-ID per pool and then emit all transaction groups higher than the stored value. Each metric should then use the transaction birth time (in combination with the system boot time) as metric timestamp, contain the pool name as tag and all transaction statistic columns as fields. Restricting this to only complete (state is c) transactions is fine.

What do you think? If you want to sample those transactions groups you can still do this using processors or aggregators...