Support distributed `ANALYZE`

[waynexia]

What problem does the new feature solve?

EXPLAIN ANALYZE only contains the execution result in frontend and the gRPC time to each datanodes. It works but can be more detailed.

What does the feature do?

Also show the detailed ANALYZE result in each datanode. I.e., implement the distributed ANALZYE plan

Implementation challenges

No response

[NiwakaDev]

@waynexia

What kind of output do you expect?

For example, something like:

+----------------------------------------------------+
| component | plan_type         | plan               |
+----------------------------------------------------+
| frontend  | Plan with Metrics | RepartitionExec    |
+----------------------------------------------------+
| node1     | ~                 | ~                  |
+----------------------------------------------------+
| node2     | ~                 | ~                  |
+----------------------------------------------------+
| node3     | ~                 | ~                  |
+----------------------------------------------------+

[waynexia]

Considering the plan might be further distributed into more parts, I'd prefer to use a tuple (stage, node) to distinguish each part. At present we only have two stages, the first is executed in datanode and the second is executed in frontend. For your example, it would become something like the following.

+------------------------------------------------------------------------+
| stage   | node                | plan_type         | plan               |
+------------------------------------------------------------------------+
| stage 2 | node 1 (addr: ...)  | Plan with Metrics | RepartitionExec    |
+------------------------------------------------------------------------+
| stage 1 | node 1 (addr: ...)  | ~                 | ~                  |
+------------------------------------------------------------------------+
| stage 1 | node 2 (addr: ...)  | ~                 | ~                  |
+------------------------------------------------------------------------+
| stage 1 | node 3 (addr: ...)  | ~                 | ~                  |
+------------------------------------------------------------------------+

Some key point:

• Node number are in the scope of stage. Start from 1 in each stage
• Stage is counted bottom-up. The first evaluated is stage 1.
• Maybe we don't need to assign node number to each node, the addr might be enough.

[NiwakaDev]

@waynexia

I'd prefer to use a tuple (stage, node) to distinguish each part.

I see. By the way, if the verbose option is set, I guess that the output is as follows:

+--------------------------------------------------------------------------------------------
| stage   | node                | plan_type         | plan               | output rows | ...
+--------------------------------------------------------------------------------------------
| stage 2 | node 1 (addr: ...)  | Plan with Metrics | RepartitionExec    | 2           | ...
+--------------------------------------------------------------------------------------------
| stage 1 | node 1 (addr: ...)  | ~                 | ~                  | 0           | ...
+--------------------------------------------------------------------------------------------
| stage 1 | node 2 (addr: ...)  | ~                 | ~                  | 1           | ...
+--------------------------------------------------------------------------------------------
| stage 1 | node 3 (addr: ...)  | ~                 | ~                  | 1           | ...
+--------------------------------------------------------------------------------------------

arrow-datafusion generates some values like output rows as the plan_type column, but in our use-case I guess it is better for the output to be format like the immediately above example in order to avoid duplication of (stage, node).

[NiwakaDev]

@waynexia

If you agree with the above format, I would like to work on this issue.

[waynexia]

Ahh, sorry for the delay.

Other parts look good to me. But things like output_rows are execution statistics, which might be missing. A simple workaround is to put all those statistics into a single string, if necessary.

[waynexia]

Hi @NiwakaDev, just a friendly ping. Do you have an initial plan or a rough structure for this? I wondered if you would like to have a discussion on any undetermined thing or question.

[NiwakaDev]

@waynexia

Do you have an initial plan or a rough structure for this? I wondered if you would like to have a discussion on any undetermined thing or question.

Here's an initial plan:

1. Send DistributedAnalyzePlan (Custom Logical plan) to each datanode.
   • DistributedAnalyzePlan simply wraps input_stream like: https://github.com/apache/arrow-datafusion/blob/main/datafusion/expr/src/logical_plan/plan.rs#L2211-L2218.
2. Execute DistributedAnalyzeExec on each datanode side like https://github.com/apache/arrow-datafusion/blob/main/datafusion/physical-plan/src/analyze.rs#L143-L247.
3. Integrate each result on the frontend side like MergeScan.

DistributedAnalyzeExec outputs:
The difference from the normal AnalyzeExec is that there're two types of output.

two types of output:

1. input_stream.next()
2. datanode analyze result like AnalyzeExec (https://github.com/apache/arrow-datafusion/blob/main/datafusion/physical-plan/src/analyze.rs#L201-L247)

While AnalyzeExec ignores input_stream.next(), in Distributed AnalyzeExec, I guess frontend needs both outputs to construct the above result (#2374 (comment)).

I haven't yet come up with a solution on how to send both input_stream.next() and output like AnalyzeExec from each datanode to frontend.

[waynexia]

Thanks for your thoughtful investigation 👍

I have one concern about passing intermediate metrics (those rendered in AnalyzeExec), which are transformed into string literal, and become hard to reuse in later phases like aggregate or filter.

Thus I came up with another way: encode and transfer metrics (MetricsSet in datafusion, specifically) along with data for each query. And report it to the user on demand, just like how an AnalyzeExec works in a single instance, drop data and execute report per-plan metrics.

For transferring metrics together with data, I've submitted a PR to add corresponding fields on proto file GreptimeTeam/greptime-proto#130 (if we decide to go this way, we can define some general metric in the proto message, instead of string-string map)

Then we don't need DistributedAnalyzePlan for datanodes. But only an uppermost, customized AnalyzeExec to extract and render the "distributed" metrics.

[waynexia]

One thing I haven't figured out is how we handle different metrics from different nodes. In datafusion, those metrics are attached to the plan itself. E.g., a join plan has two children, and each child can keep its own metrics. But here we don't have actual children nodes beside MergeScan. So how can we keep the tree structures for metrics from them?

If we don't need to distinguish metrics from each node, we can aggregate them into one sub-tree in MergeScan, since they are going to have the same physical plan. But if what we want is above per-phase and per-node analyze, this can't help.

[NiwakaDev]

@waynexia

One thing I haven't figured out is how we handle different metrics from different nodes. In datafusion, those metrics are attached to the plan itself. E.g., a join plan has two children, and each child can keep its own metrics. But here we don't have actual children nodes beside MergeScan. So how can we keep the tree structures for metrics from them?

The issue you wrote is related to the below code? Sorry, I might be wrong because I'm not familiar with datafusion.
https://github.com/apache/arrow-datafusion/blob/main/datafusion/physical-plan/src/display.rs#L243-L253
https://github.com/apache/arrow-datafusion/blob/main/datafusion/physical-plan/src/visitor.rs#L24-L34

Thus I came up with another way: encode and transfer metrics (MetricsSet in datafusion, specifically) along with data for each query. And report it to the user on demand, just like how an AnalyzeExec works in a single instance, drop data and execute report per-plan metrics.

If we don't need analyze per-node, I agree with this.

[NiwakaDev]

@waynexia

By the way, if we implement your idea, what kind of output do you expect?

[waynexia]

One thing I haven't figured out is how we handle different metrics from different nodes. In datafusion, those metrics are attached to the plan itself. E.g., a join plan has two children, and each child can keep its own metrics. But here we don't have actual children nodes beside MergeScan. So how can we keep the tree structures for metrics from them?

The issue you wrote is related to the below code? Sorry, I might be wrong because I'm not familiar with datafusion.

greptimedb/src/query/src/dist_plan/merge_scan.rs

Lines 264 to 266 in c7b3677

	fn children(&self) -> Vec<Arc<dyn ExecutionPlan>> {
	vec![]
	}

MergeScanExec is the leaf node in Frontend's execution plan tree. It submits a part of the query to datanodes and merges result from them (hence "merge scan"). Since those part of plan is not executed in frontend, it neither has children nor their metrics. This prevents the visitor method you mentioned above from walking into MergeScanExec.

We can retrieve metrics from datanodes, but I'm afraid we have to keep and access those "remote metrics" in a different way for this reason.

By the way, if we implement your idea, what kind of output do you expect?

I would like to have two forms. One is distinguished with the tuple (phase, node), we can retrieve un-aggregated, per-phase and per-node metrics for detailed analysis. Another may look closer to the ordinary ANALYZE, which hides the distributed execution's details and gives a rough and aggregated result.

[waynexia]

Update: at this stage, it is clear that we have to find a way to pass data and execution metrics together in the same query call. @shuiyisong and I are trying to add a method to SendableRecordBatchStream to provide the corresponding execution metrics. But it's still undetermined how to define, organize and expose those metrics.

I haven't yet come up with a solution on how to send both input_stream.next() and output like AnalyzeExec from each datanode to frontend.

We can assume this issue is resolved (if everything works as expected...) and bring this ticket forward 🙌 @NiwakaDev