feat(sink): decouple starrocks commit from risingwave commit #16816

ly9chee · 2024-05-19T15:23:57Z

I hereby agree to the terms of the RisingWave Labs, Inc. Contributor License Agreement.

What's changed and what's your intention?

The current StarRocks sink implementation using StreamLoad API writes data to StarRocks. It works well when the upstream does not change frequently. However, if the upstream changes frequently, many versions of the table being written to are generated, which will impact the query performance and bring significant pressure to the BE Compactor.

Implementation details

This PR attempts to solve this problem by leveraging the Starrocks Stream Load Transaction Interface. The idea is inspired by RFC Support sink coordinator and RFC Decouple iceberg sink commit from risingwave checkpoint. Despite these two RFCs being designed for data lakes, but after doing some research, I found OLAP System like StarRocks can apply the same pattern as described in these RFCs as well, i.e., Changing the current StreamLoad to the Stream Load Transaction Interface.

If you run a load job by using a program, the Stream Load transaction interface allows you to merge multiple mini-batches of data on demand and then send them all at once within one transaction by calling the /api/transaction/commit operation. As such, fewer data versions need to be loaded, and load performance is improved.

As the doc says, after a transaction is created, we can write multiple mini-batches to the transaction through /api/transaction/load interface. StarRocks will take care of merging these batches into a larger one when the /api/transaction/commit is called. So in SinkWriter implementation, we can write data to the transaction at every barrier, but only commit the transaction when a checkpoint barrier is received, which is delayed by commit_checkpoint_interval. In fact, the commit is requested after the coordinator collects the metadata from each SinkWriter.

Assume the barrier interval is 1 second and the checkpoint frequency is 1. The current sink implementation will commit data to StarRocks at every barrier, thus producing a version every second. But this PR will commit data to StarRocks only when the user-specified commit_checkpoint_interval is reached. If we set commit_checkpoint_interval to 10, a version is created roughly every 10 seconds.

Minor refactors

URL encoding is applied to the username and password during the construction of mysql_conn_uri to allow special characters to be included.
StarrocksCommon.partial_update option has been moved up to StarrocksConfig, which is a sink-only option.

Simple benchmark on my local environment

Sinking 5 million rows to StarRocks

Current implementation, parallelism=4

This PR, parallelism=4, commit_checkpoint_interval=10

We can see that the Version has reduced from 352 to 45

Notes

Compare to the current sink implementation, there is some slight overhead within a commit_checkpoint_interval period due to the begin, prepare and commit transaction meta requests being made.
The duration of commit_checkpoint_interval should not exceed FE's config parameter stream_load_default_timeout_second; otherwise, the transaction will timeout and the sink will fail repeatedly.
This PR should be compatible with the current implementation, because the current implementation is just a special case of setting commit_checkpoint_interval to 1, which is the default.
The transaction interface was introduced in StarRocks 2.4, but the current StarRocks sink seems only work with StarRocks version >= 2.5 due to the usage of information_schema.tables_config, which was introduced in version 2.5.
Even if we can reduce the versions by setting commit_checkpoint_interval to a higher value, we sacrifice the freshness. BTW, we can further reduce the versions by lowering the sink parallelism as well.

This PR's code is base on #16777, because we all used DecoupleCheckpointLogSinkerOf.

Checklist

I have written necessary rustdoc comments
I have added necessary unit tests and integration tests
I have added test labels as necessary. See details.
I have added fuzzing tests or opened an issue to track them. (Optional, recommended for new SQL features Sqlsmith: Sql feature generation #7934).
My PR contains breaking changes. (If it deprecates some features, please create a tracking issue to remove them in the future).
All checks passed in ./risedev check (or alias, ./risedev c)
My PR changes performance-critical code. (Please run macro/micro-benchmarks and show the results.)

My PR contains critical fixes that are necessary to be merged into the latest release. (Please check out the details)

Documentation

My PR needs documentation updates. (Please use the Release note section below to summarize the impact on users)

Release note

If this PR includes changes that directly affect users or other significant modifications relevant to the community, kindly draft a release note to provide a concise summary of these changes. Please prioritize highlighting the impact these changes will have on users.

src/connector/src/sink/deltalake.rs

Co-authored-by: TennyZhuang <zty0826@gmail.com>

fuyufjh

Generally LGTM

src/connector/src/sink/deltalake.rs

fuyufjh · 2024-05-21T04:27:28Z

src/connector/src/sink/doris_starrocks_connector.rs

+            let mut parsed_be_url =
+                Url::parse(&be_url).map_err(|err| SinkError::DorisStarrocksConnect(err.into()))?;
+
+            if fe_host != LOCALHOST && fe_host != LOCALHOST_IP {


Hmm, actually any IP of 127. refers to localhost i.e. loop-back. Does SR guarantee to return 127.0.0.1?

This snippet is actually borrowed from previous version 😅。I think it may related to solve some starrocks cluster network issues that deployed for testing, which may redirect BE to a localhost address.

risingwave/src/connector/src/sink/doris_starrocks_connector.rs

Lines 235 to 241 in 5dc3861

if be_host == LOCALHOST || be_host == LOCALHOST_IP {

// if be host is 127.0.0.1, we may can't connect to it directly,

// so replace it with fe host

parsed_be_url

.set_host(Some(self.fe_host.as_str()))

.map_err(|err| SinkError::DorisStarrocksConnect(err.into()))?;

be_url = parsed_be_url.as_str().into();

Related PR #16018

fuyufjh · 2024-05-21T04:31:54Z

src/connector/src/sink/doris_starrocks_connector.rs

+        let be_url = try_get_be_url(&resp, self.fe_host.clone())?.ok_or_else(|| {
+            SinkError::DorisStarrocksConnect(anyhow::anyhow!("Can't get doris BE url",))
+        })?;


Should use the error returned by try_get_be_url, rather than anyhow!

Suggested change

let be_url = try_get_be_url(&resp, self.fe_host.clone())?.ok_or_else(|| {

SinkError::DorisStarrocksConnect(anyhow::anyhow!("Can't get doris BE url",))

})?;

let be_url = try_get_be_url(&resp, self.fe_host.clone())?;

~~You're right, the returned error should not be wrapped twice.~~
try_get_be_url will return a Result<Option<Url>>, returning Ok(None) indicates this request does not redirect. For exmaple, /api/transaction/commit interface sometimes will not redirect to BE. In this case, commit request should be treated as finished. But other interfaces /api/transaction/begin, /api/transaction/load and /api/db/table/_stream_load (now only used in Doris sink) are expected redirect to BE. ok_or_else is used to handle this case where no redirection is considered an error.

And I don't find any documents in starrocks that describe BE redirection behavior. we cannot make sure in which situation the FE should redirect to BE or not.

BTW, if reqwest supports auto redirection without removing sensitive headers like Authorization, we may make our code more elegant.

fuyufjh · 2024-05-21T04:32:32Z

src/connector/src/sink/doris_starrocks_connector.rs

+        let be_url = try_get_be_url(&resp, self.fe_host.clone())?.ok_or_else(|| {
+            SinkError::DorisStarrocksConnect(anyhow::anyhow!("Can't get doris BE url",))
+        })?;


Ditto.

Suggested change

let be_url = try_get_be_url(&resp, self.fe_host.clone())?.ok_or_else(|| {

SinkError::DorisStarrocksConnect(anyhow::anyhow!("Can't get doris BE url",))

})?;

let be_url = try_get_be_url(&resp, self.fe_host.clone())?;

By the way, why not do try_get_be_url() when the connector initialize and use the URL in later?

This is because we send requests to FE, FE acts like a balancer and will redirect the request to a BE based on its internal balance algorithm, so we need to get the BE url for each request.

fuyufjh · 2024-05-21T04:34:06Z

src/connector/src/sink/doris_starrocks_connector.rs

+                    String::from_utf8(raw)
+                        .map_err(|err| SinkError::DorisStarrocksConnect(err.into()))?


Internal err was lost. Please either pass the err as source (recommended) or use err.as_report() to get the proper error message.

See also https://www.notion.so/risingwave-labs/A-Guide-to-Error-Handling-that-Just-Works-Part-I-e108f8d0845b4bddaa0e5843791a8cac

I just send an access request for this doc, please check.

The owner of the doc is vacationing. Let me export it.
A_Guide_to_Error_Handling_that_Just_Works_(Part_I).pdf

I have replaced SinkError::DorisStarrocksConnect(err.into()) with SinkError::DorisStarrocksConnect(anyhow!(err)) to keep the uniformity with other sinks.

src/connector/src/sink/doris_starrocks_connector.rs

src/connector/src/sink/decouple_checkpoint_log_sink.rs

src/connector/src/sink/deltalake.rs

ly9chee · 2024-05-21T08:53:36Z

Should we run doris sink e2e test as well?, There're some refactors made to doris_starrocks_connector::InserterInnerBuilder @xxhZs

fuyufjh

LGTM!

…ization issue

…chee/risingwave into decouple-starrocks-commit-v2

xxhZs and others added 9 commits May 16, 2024 15:52

add

23796a2

fix ut

11305fc

Merge branch 'main' into xxh/add-decouple-deltalake

aa1bbe6

fix ci

902de74

Merge branch 'risingwavelabs:main' into add-decouple-deltalake

64a3b84

implement decouping logic

2683651

fix commit logic

64da9a4

fix rustfmt

752a770

add with_options

a5601b4

StrikeW requested review from fuyufjh, ZENOTME and xxhZs May 20, 2024 02:40

ly9chee added 3 commits May 20, 2024 11:18

separate doris/starrocks success status

cfc4642

reorganize code

5daac28

strip Arc from reqwest::Client

3168941

TennyZhuang reviewed May 20, 2024

View reviewed changes

src/connector/src/sink/deltalake.rs Outdated Show resolved Hide resolved

fix comment

73cc5dd

Co-authored-by: TennyZhuang <zty0826@gmail.com>

fuyufjh reviewed May 21, 2024

View reviewed changes

xxhZs added the ci/run-e2e-starrocks-sink-tests label May 21, 2024

ly9chee added 3 commits May 21, 2024 15:33

add rustdoc for DecoupleCheckpointLogSinkerOf

42d5b6d

resolve conflicts

d360823

refine error handling

e35d6c1

xxhZs added the ci/run-e2e-doris-sink-tests label May 21, 2024

fuyufjh approved these changes May 21, 2024

View reviewed changes

hzxa21 added the user-facing-changes Contains changes that are visible to users label May 22, 2024

ly9chee added 3 commits May 24, 2024 19:16

Merge branch 'risingwavelabs:main' into decouple-starrocks-commit-v2

55e0a6c

add missing with_options comment and fix stream_load_timeout deserial…

d48599b

…ization issue

Merge branch 'decouple-starrocks-commit-v2' of https://github.com/ly9…

b5deec6

…chee/risingwave into decouple-starrocks-commit-v2

xxhZs approved these changes May 29, 2024

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(sink): decouple starrocks commit from risingwave commit #16816

feat(sink): decouple starrocks commit from risingwave commit #16816

ly9chee commented May 19, 2024 •

edited

fuyufjh left a comment

fuyufjh May 21, 2024

ly9chee May 21, 2024 •

edited

xxhZs May 21, 2024

fuyufjh May 21, 2024

ly9chee May 21, 2024 •

edited

ly9chee May 21, 2024

ly9chee May 21, 2024

fuyufjh May 21, 2024

fuyufjh May 21, 2024

ly9chee May 21, 2024

fuyufjh May 21, 2024

ly9chee May 21, 2024

fuyufjh May 21, 2024

ly9chee May 21, 2024

ly9chee commented May 21, 2024

fuyufjh left a comment

	if be_host == LOCALHOST \|\| be_host == LOCALHOST_IP {
	// if be host is 127.0.0.1, we may can't connect to it directly,
	// so replace it with fe host
	parsed_be_url
	.set_host(Some(self.fe_host.as_str()))
	.map_err(\|err\| SinkError::DorisStarrocksConnect(err.into()))?;
	be_url = parsed_be_url.as_str().into();

		String::from_utf8(raw)
		.map_err(\|err\| SinkError::DorisStarrocksConnect(err.into()))?

feat(sink): decouple starrocks commit from risingwave commit #16816

Are you sure you want to change the base?

feat(sink): decouple starrocks commit from risingwave commit #16816

Conversation

ly9chee commented May 19, 2024 • edited

What's changed and what's your intention?

Implementation details

Minor refactors

Simple benchmark on my local environment

Notes

Checklist

Documentation

Release note

fuyufjh left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ly9chee May 21, 2024 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ly9chee May 21, 2024 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ly9chee commented May 21, 2024

fuyufjh left a comment

Choose a reason for hiding this comment

ly9chee commented May 19, 2024 •

edited

ly9chee May 21, 2024 •

edited

ly9chee May 21, 2024 •

edited