Compute Introspection sources/tables/views improvements

Summary

We have identified several issues with the sources documented here. In particular, some key views are missing (e.g., a way to connect channel IDs to the operators they connect), and there are various existing ones that have inconsistently named fields or inconsistent semantics. Furthermore, the documentation of the various relations is incomplete (e.g., it doesn't make clear whether each relation is per-replica or global).

Goals

Compute introspection relations should be useful and easy to understand for internal debugging.
They should be useful and easy to use for building external introspection UI.

Non-Goals

The relations do not need to be polished enough for external consumption (although it's hoped they prove useful, we don't consider it a bug if they're not). Customer-friendly introspection will be provided in the web UI by the "external introspection" project.
We make no attempt to maintain backwards compatibility with previously-existing queries, although as part of this work we will update the internal tools that use the introspection relations.

Description

Schema inconsistencies and inelegancies

In this section, we describe the various issues that make the schema confusing.

Per-worker vs. aggregated

Several relations have a worker_id column. Technically, the semantics of this column are the same across all the relations: it corresponds to the worker on which the data was measured. However, in practice the behavior is very different, and falls into a few separate categories, as described below.

Sharded relations

Example: mz_arrangement_sizes. These are relations that report some value that is expected to differ across all workers, such that the semantically meaningful value is (often) the sum of all the per-worker values. For example, a user interested in the total size of each arrangement would normally want to issue a query like SELECT operator_id, sum(records), sum(batches) FROM mz_arrangement_sizes GROUP BY operator_id.

There are also a few relations where sum is not the most interesting aggregate; for example, the min across workers in mz_compute_frontiers. "Sharded" might not be the most precise way to describe these, but I include them in this section because the point is basically the same.

The existence of the per-worker underlying sources is not completely useless, e.g. because they are useful for detecting skew. However, they are probably less commonly useful than a summed view would be. Unforunately, this summed view does not exist for most such relations (one counterexample is mz_records_per_dataflow and mz_records_per_dataflow_global).

Duplicated relations

Example: mz_dataflow_channels. These relations report static structural properties of dataflows. As such, they will always report exactly the same values across all workers, except in one of two cases:

Our dataflow-rendering logic is buggy, and we invalidly produce differently-shaped dataflows on different workers, or
Due to a race condition, where the relation is queried just after one worker has emitted a log message corresponding to the dataflow, but just before another worker has.

These cases are unlikely to be observable in practice, the first because dataflow mismatches will usually cause unpredictable behavior (including crashes) long before we have a chance to usefully investigate them via introspection sources, and the second because such race conditions will typically be very short-lived.

Thus, the argument for making global views (e.g., that simply filter the underlying sources by worker 0) is even stronger than for the "Sharded relations" case.

`_internal` and `raw` variants

Several of the relations are views over very similar underlying sources whose names are suffixed with _internal or infixed with raw, which are similar enough to be possibly confusing, but mean something completely different. _internal relations are those that count (or sum) something in the Differential Dataflow count field; the corresponding non-_internal views are simple aggregates that move that count into the data. For example, mz_arrangement_sharing is defined as

SELECT
    operator_id,
    worker_id,
    pg_catalog.count(*) AS count
FROM mz_internal.mz_arrangement_sharing_internal
GROUP BY operator_id, worker_id"

This dimension is applied inconsistently: not all relations that report a count or sum in the data are derived from a corresponding _internal source.

raw variants, on the other hand, measure some time value in nanoseconds, whereas their corresponding non-raw views convert the times to the SQL interval type.

Confusingly, not all times are reported as intervals; for example, mz_scheduling_parks has slept_for and requested columns, whose type is bigint and whose unit is undocumented (in fact, it is nanoseconds).

The trade-off between intervals and nanoseconds timestamps is that the former is more ergonomic but also has only microsecond precision.

Inconsistent presence of `sum` column in histograms

Some histograms include a per-bucket sum, which allows consumers to compute average or total values precisely. Confusingly, not all histograms report such a column (e.g. mz_worker_compute_delays does not, whereas mz_worker_raw_compute_delays does).

We introduced the sum column initially to support Prometheus histograms, which expects to know about cumulative sums and cannot be convinced otherwise. However, it is possible to provide approximate sums, based on multiplying the bucket values with the bucket counts.

Lack of clarity around per-replica vs. global relations

Many of the relations are logging sources that run locally in clusters, are not persisted, and report different data depending on the currently selected replica. Views based on these sources are also per-replica. It is not straightforward to tell from our docs which is the case for each relation.

Furthermore, sometimes the same logical value is reported by two relations, one reporting clusterd's view of the world, and the other reporting environmentd's. For example, mz_compute_frontiers vs. mz_cluster_replica_frontiers. However, the names do not make it clear what the distinction is.

Suggested fixes for schema inconsistencies

All per-worker relations should have a corresponding "global" view where the results are aggregated across workers in the way most likely to be useful (for example, summing, taking the minimum, or restricting to worker 0). Because these global views are more likely to be useful than the underlying sources, they should have the basic name (e.g., mz_arrangement_sharing). The underlying source should have a name with per_worker suffixed (e.g., mz_arrangement_sharing_per_worker). Furthermore, the per-worker variant should be clearly less prominent in documentation (either at the bottom of the page, or in a "see more" section that has to be clicked to expand).
The expectations about per-worker relations should be clearly documented (e.g., we should clearly say whether the values for each worker are expected to be the same during normal operation).
raw variants should be removed, and all introspection relations should present durations as nanoseconds timestamps. This allows us to provide the full precision, while also allowing users to cast the durations to intervals if they find those more convenient.
All relations that report a simple count or sum should be views based on a *_raw variant that maintains the count or sum in the Differential diff field. These relations previously had an _internal suffix, but _raw is less likely to be confused with the mz_internal schema, and better reflects that fact that these relations contain unaggregated data.
All columns containing nanoseconds timestamps should have names suffixed with _ns.
The sum column should be removed from all histogram sources. Maintaining that field has some cost, in the form of additional state we must keep, a roughly doubled amount of updates, and code complexity. On the other hand, it is not clear how much value the sum field provides given that we can approximate it.
All histogram views or sources should be suffixed with _histogram, to make their role more obvious.

Possibly useful missing views

Channel-to-operator mapping

See GitHub PR.

Scope child to parent mapping

The only way to understand the structure of scope nesting currently is to use the mz_dataflow_addresses relation and do some tedious list manipulation. A simple view of (id, parent_id) would make this much simpler.

Transitive object dependencies and end-to-end lag

We should add the transitive closure of mz_object_dependencies, called e.g. mz_transitive_object_dependencies.

This will allow us to see the end-to-end frontier lag for any dataflow.

More views enriched with operator names, dataflow names, etc.

(Not quite sure about this -- it's tricky to balance ease-of-use with not blowing up the set of documented relations)

Alternatives

None known to me

Open questions

Are there further views that could be useful but that we don't have access to? We should fill this out as more become known to us (e.g., while debugging customer issues)
Is there a better name than _internal for the non-aggregated sources?

20230225_compute_introspection_schema.md 9.5 KB Permalink History Raw