Add metadata to push payload #9694

salvacorts · 2023-06-13T07:50:25Z

What this PR does / why we need it:
We are adding support for attaching labels to each log line. This is one of the series of the PRs broken up to make it easier to review changes.

This PR updates the push payload to send labels with each log entry optionally. The log labels are supposed to be in the same format as the stream labels. Just to put it out, here is how it would look for proto and json push payload with same data:

proto(endpoint: (/loki/api/v1/push|/api/prom/push), Content-Type: application/x-protobuf)(payload built using push.Stream):

push.Stream{
	Entries: []logproto.Entry{
			{
				Timestamp: time.Unix(0, 1688515200000000000),
				Line:      "log line",
				Labels:    `{foo="bar"}`,
			},
	},
	Labels: `{app="test"}`,
}

v1(endpoint: /loki/api/v1/push, Content-Type: application/json):

{
    "streams": [{
        "stream": {
            "app": "test"
        },
        "values": [
            ["1688515200000000000", "log line", {
                "foo": "bar"
            }]
        ]
    }]
}

legacy-json(/api/prom/push, Content-Type: application/json):

{
    "streams": [{
        "labels": "{app=\"test\"}",
        "entries": [{
                "ts": "2023-07-05T00:00:00.000000000Z",
                "line": "log line",
                "labels": "{foo=\"bar\"}"
            }]
    }]
}

Which issue(s) this PR fixes:

Special notes for your reviewer:

We may need to add more thoughtful tests.

Checklist

Reviewed the CONTRIBUTING.md guide (required)
Documentation added
Tests updated
CHANGELOG.md updated
Changes that require user attention or interaction to upgrade are documented in docs/sources/upgrading/_index.md
For Helm chart changes bump the Helm chart version in production/helm/loki/Chart.yaml and update production/helm/loki/CHANGELOG.md and production/helm/loki/README.md. Example PR

pkg/logql/log/fmt_test.go

integration/client/client.go

…n push request

owen-d

In the future, please add a description about the PR to help reviewers. A bit of history + the design choices taken and the implications is sufficient and makes it much easier to know what I'm looking at.

As for the PR, I think JSON content types should support json sub-structure like we do for stream labels in https://grafana.com/docs/loki/latest/api/#push-log-entries-to-loki

Basically:

{
  "streams": [
    {
      "stream": {
        "label": "value"
      },
      "values": [
          [ "<unix epoch in nanoseconds>", "<log line>", {"traceID": "foo"} ],
          [ "<unix epoch in nanoseconds>", "<log line>", {"order_id": "bar"} ]
      ]
    }
  ]
}

owen-d · 2023-06-22T18:35:21Z

pkg/loghttp/query.go

 type PushRequest struct {
-	Streams []*Stream `json:"streams"`
+	Streams []LogProtoStream `json:"streams"`


Previously this stored a pointer, now it stores a struct. Not sure we want this (would need to be tested independently for performance)

This comes from this PR:
#1145

The only reason why I think this is a pointer is so we don't make a copy of the stream when calling NewStream which receives a ptr to loghttp.Stream
https://github.com/grafana/loki/pull/1145/files#diff-71619e1c80a73b34eade235a55d012a0ddbb3375b8d4ac89c1f4fd672145b915R34

With Sandeep's changes, we no longer use that since now we decode from json to loghttp.PushRequest to then cast it to logproto.PushRequest

loki/pkg/util/unmarshal/unmarshal.go

Line 22 in 44b0932

Streams: *(*[]logproto.Stream)(unsafe.Pointer(&request.Streams)),

So having that as a ptr or not shouldn't make much difference.

owen-d · 2023-06-22T18:36:04Z

pkg/loghttp/query.go

+
+// LogProtoStream helps with unmarshalling of each log stream for push request.
+// This might look un-necessary but without it the CPU usage in benchmarks was increasing by ~25% :shrug:
+type LogProtoStream struct {


I think it makes more sense to implement UnmarshalJSON on logproto.Stream directly -- is that feasible?

~~I agree!~~

Edit: I think that's not possible right-away. logproto.Stream is an alias to push.Stream provided in the package github.com/grafana/loki/pkg/push which is used in this package.

UnmarshalJSON uses the LabelsSet, so pkg/push would need to include pkg/loki which would end up in a cycling dependency.
https://github.com/grafana/loki/pull/9694/files#diff-5b050fb13a302741e2f4a781fa54987e96da67696ae36ea41aa971ef431bfeccR84

I think this Unmarshalling method make sense to live here since:

This pkg is the one needing to populate a logproto.Stream from a json. Moreover, this json has a format for entries (json array) that only this package should know/care about.

Workarounding the cycling dependency would require adding more changes than the ones required to implement it here.

I think we can bring all the parsing code to a single place when we work on the endpoint for supporting native OTLP ingestion.

owen-d · 2023-06-22T19:12:20Z

pkg/loghttp/entry.go

@@ -18,6 +18,7 @@ func init() {
 type Entry struct {
 	Timestamp time.Time
 	Line      string
+	Labels    string


Making this a string feels counterintuitive compared to using map[string]string, which is what we use via the LabelSet struct used in our v1 push payloads in loghttp/push. Point being, I think we should use JSON directly (not json with embedded stringified json fields) when application/json is the Content-Type. When snappy-encoding is used, logproto can use the string variants for consistency with how it treats labels.

The problem I see with making this a map, and keeping as a string the labels for an entry in logproto, is that converting from loghttp.Entry to logproto.Entry is going to be way slower.

Right now (both with string labels), we just cast the pointer for the entries array of a stream since they have the same memory layout.

loki/pkg/loghttp/query.go

Line 236 in 44b0932

entries := *(*[]logproto.Entry)(unsafe.Pointer(&s.Entries))

If we decide to make one a map and another one a string, we won't be able to do that cast anymore (different memory layouts). Hence, we'll need to go entry by entry converting it from loghttp.Entry to logproto.Entry.

IMO, we should make both use the same layout.

I implemented this in f43dfab

This is what we need to do instead of the cast:

loki/pkg/loghttp/query.go

Lines 236 to 239 in f43dfab

entries := make([]logproto.Entry, len(s.Entries), len(s.Entries))

for i, e := range s.Entries {

entries[i] = e.ToProto()

}

If we decide to keep this as a string, we can revert the commit. If we decide to make logproto.Entry to use a map for labels, we can build on top of it.

according to our API docs, we use JSON object for stream labels )https://grafana.com/docs/loki/latest/api/#push-log-entries-to-loki
so, it makes sense to use JSON object for line metadata also… but you are right @salvacorts that there is no need to convert it to the string when we convert loghttp.Entry to logproto.Entry ….

@salvacorts You're right, making this anything but a string here will result in performance losses because this it would need to parse labels potentially on every entry coming back from subqueries. Since this happens before we filter out unnecessary entries (over limit, etc), this can quickly result in higher resource utilization, OOMs, etc.

Now, we have a few options:

Keep this as a string in memory, but turn it into a map when return the json response from the http api. This will maintain performance (no conversion from protos) & correctly return a map via the json api.

Use LabelSet for this field (stored as a map). This requires parsing strings -> json for potentially every log line returned from queriers (yikes!). Even when combined with (4), this would require transforming slice -> map.

Use labels.Labels (stored as a slice). When combined with (4) this incurs no transformation cost and the resulting conversion to the json map for the external API is simple & relatively inexpensive.

Due to the performance implications here, I've changed my opinion on making EntryAdapter.labels a string for parity with StreamAdapter.labels. Instead, I think we should use the LabelAdapter pattern to send the labels in the protos as a slice and combine it with (3) to not need to pay conversion costs.

Does this make sense?

P.S. We'll actually need to do this on the Entry and not EntryAdapter types (we only keep the adapter ones around for bench comparisons, but they're unused). Happy to talk through this more when I've got a bit more time, but we did this primarily for serialization optimizations.

I would vote for 3 and 4. looks like the most optimal way for us. wdyt @salvacorts ?

Agree. Implemented it in 4028758. We now use labels.Labels in both loghttp.Entry and logproto.Entry.

I wonder if we should take care of duplicated labels. I think may not (?) provided the push payload in json is an object so it should have duplicated keys (is that somehow enforced?).

Note that proto3 supports maps. In the wire, the map is the same as an array so it needs extra logic to have the map API. I ran a benchmark and looks like maps compared to arrays are ~x3 slower to marshal and ~x2 slower to unmarshal (happy to share/discuss the benchmark). Therefore I agree the way to go is using arrays in protobuf.

vlad-diachenko

Overall looks great )

pkg/push/push.proto

integration/client/client.go

vendor/github.com/grafana/loki/pkg/push/push.proto

owen-d · 2023-06-23T20:42:02Z

pkg/loghttp/query.go

+	if parseError != nil {
+		return nil, parseError
+	}
+	if err != nil {


This feels weird -- is there a chance err can be non-nil but parseError is nil?

The error handling of the implementation for jsonparser.ArrayEach doesn't seems correct to me. The error being passed to the callback is weird as it'll always be nil. Then, after calling the callback, the library checks for the value of the error, which will always be nil since it's passed by value to the callback. There's an issue open for this:
buger/jsonparser#255

IIUC, parseError can be nil while err can be non-nil as it can return for example a MalformedJsonError.
Having said that, if err is not nil, we should return err, not parseError.

Done with 384d556

owen-d · 2023-06-23T20:45:23Z

pkg/loghttp/query.go

+		case 2: // labels
+			labels := make(LabelSet)
+			err := jsonparser.ObjectEach(value, func(key, val []byte, _ jsonparser.ValueType, _ int) error {
+				labels[yoloString(key)] = yoloString(val)


Don't we need to check types for the values here to ensure they're strings?

Done with 384d556

pkg/push/push.proto

owen-d · 2023-06-23T21:09:01Z

pkg/loghttp/entry.go

@@ -126,6 +143,8 @@ func (EntryEncoder) Encode(ptr unsafe.Pointer, stream *jsoniter.Stream) {
 	stream.WriteRaw(`"`)
 	stream.WriteMore()
 	stream.WriteStringWithHTMLEscaped(e.Line)
+	stream.WriteMore()
+	stream.WriteString(e.Labels)


These types are used in the external API, meaning we'd return stringified labels as a field, which we shouldn't do. We should use json semantics here. Ideally, we should also reflexively be able to encode -> decode, meaning json unmarshaling should expect a json object of non-indexed labels as well.

Done with f43dfab

vendor/github.com/grafana/loki/pkg/push/types.go

vlad-diachenko

LGTM

vlad-diachenko · 2023-07-12T10:59:35Z

pkg/loghttp/entry.go

+					Name:  yoloString(key),
+					Value: yoloString(value),


I have doubts about using yoloString here.
if underlying array of value is reused , the values here will be also updated

I think we should remove yoloString here. I had used it earlier because the labels were anyways converted to string immediately. The references to string are now being retained longer so it is not safe to use yoloString here.

As discussed in Slack, removed.

vlad-diachenko · 2023-07-12T11:04:29Z

pkg/loghttp/query.go

+func yoloString(buf []byte) string {
+	return *((*string)(unsafe.Pointer(&buf)))
+}


the scariest thing in the PR ) we should use it only if we check that the buffer is not updated, however, a lot of functions in stack trace mean that somebody can do some optimization to reuse the slice and it will break all our implementation. Why do we need it here?

As discussed in Slack, removed.

Co-authored-by: Sandeep Sukhani <[email protected]>

salvacorts · 2023-07-12T14:22:31Z

I ran the changes in this PR for a couple of hours in a dev cluster and seems to be working fine:

Write path

Read path

Manually ran some queries for recent data that succeeded

pkg/loghttp/query.go

owen-d

address @sandeepsukhani 's comment about yolostring and LGTM

**What this PR does / why we need it**: In #9694, we modified the `push` pkg to import Prometheus as a dependency to use the `labels.Labels` type for the entries' non-indexed labels. Having Prometheus as a dependency is problematic since any project importing the `push` pkg will need to import Prometheus as a result. In fact, this is one of the reasons why the `push` pkg was extracted from Loki in the first place (#8259). This PR removes the dependency of Prometheus from the `push` pkg by copying some bits of the implementation for Prometheus' `labels.Labels`. We copy: - The Labels struct definition - The JSON Marshaling and Unmarshaling methods for the labels. We need this so labels are encoded as maps instead of an array of objects. --- **Notes for reviewers:** - To implement the JSON Marshaling and Unmarshaling methods the `push` pkg now depends on `golang.org/x/exp`. I think it should be fine for projects importing the `push` pkg to also depend on `golang.org/x/exp`, right?

**What this PR does / why we need it**: In #9694 we added support for adding metadata labels to each entry in a push payload. In this PR we update the following metrics to take into account the bytes needed to store those labels: - `loki_distributor_bytes_received_total` - `distributor_bytes_received` **Which issue(s) this PR fixes**: Fixes #<issue number> **Special notes for your reviewer**: **Checklist** - [ ] Reviewed the [`CONTRIBUTING.md`](https://github.com/grafana/loki/blob/main/CONTRIBUTING.md) guide (**required**) - [ ] Documentation added - [ ] Tests updated - [ ] `CHANGELOG.md` updated - [ ] If the change is worth mentioning in the release notes, add `add-to-release-notes` label - [ ] Changes that require user attention or interaction to upgrade are documented in `docs/sources/upgrading/_index.md` - [ ] For Helm chart changes bump the Helm chart version in `production/helm/loki/Chart.yaml` and update `production/helm/loki/CHANGELOG.md` and `production/helm/loki/README.md`. [Example PR](d10549e) --------- Co-authored-by: Sandeep Sukhani <[email protected]>

**What this PR does / why we need it**: In #9694, we modified the `push` pkg to import Prometheus as a dependency to use the `labels.Labels` type for the entries' non-indexed labels. Having Prometheus as a dependency is problematic since any project importing the `push` pkg will need to import Prometheus as a result. In fact, this is one of the reasons why the `push` pkg was extracted from Loki in the first place (#8259). This PR removes the dependency of Prometheus from the `push` pkg by copying some bits of the implementation for Prometheus' `labels.Labels`. We copy: - The Labels struct definition - The JSON Marshaling and Unmarshaling methods for the labels. We need this so labels are encoded as maps instead of an array of objects. --- **Notes for reviewers:** - To implement the JSON Marshaling and Unmarshaling methods the `push` pkg now depends on `golang.org/x/exp`. I think it should be fine for projects importing the `push` pkg to also depend on `golang.org/x/exp`, right? (cherry picked from commit 15af77b)

Backport 15af77b from #9937 --- **What this PR does / why we need it**: In #9694, we modified the `push` pkg to import Prometheus as a dependency to use the `labels.Labels` type for the entries' non-indexed labels. Having Prometheus as a dependency is problematic since any project importing the `push` pkg will need to import Prometheus as a result. In fact, this is one of the reasons why the `push` pkg was extracted from Loki in the first place (#8259). This PR removes the dependency of Prometheus from the `push` pkg by copying some bits of the implementation for Prometheus' `labels.Labels`. We copy: - The Labels struct definition - The JSON Marshaling and Unmarshaling methods for the labels. We need this so labels are encoded as maps instead of an array of objects. --- **Notes for reviewers:** - To implement the JSON Marshaling and Unmarshaling methods the `push` pkg now depends on `golang.org/x/exp`. I think it should be fine for projects importing the `push` pkg to also depend on `golang.org/x/exp`, right? Co-authored-by: Salva Corts <[email protected]>

**What this PR does / why we need it**: In #9694 we support adding metadata labels for each entry in the push payload. In this PR we take that metadata and add it to the entries in the chunk. Supporting serialization and deserialization of those metadata labels. --------- Signed-off-by: Vladyslav Diachenko <[email protected]> Co-authored-by: Vladyslav Diachenko <[email protected]>

**What this PR does / why we need it**: Even though the[ Loki HTTP API docs for the push endpoint][1] state that the stream label values should be strings, we previously didn't enforce this requirement. With #9694, we started enforcing this requirement, and that broke some users. In this PR we are reverting this type of assertion and adding a bunch of tests to avoid the regression in the future. [1]: https://grafana.com/docs/loki/latest/reference/api/#push-log-entries-to-loki

**What this PR does / why we need it**: Even though the[ Loki HTTP API docs for the push endpoint][1] state that the stream label values should be strings, we previously didn't enforce this requirement. With #9694, we started enforcing this requirement, and that broke some users. In this PR we are reverting this type of assertion and adding a bunch of tests to avoid the regression in the future. [1]: https://grafana.com/docs/loki/latest/reference/api/#push-log-entries-to-loki (cherry picked from commit 5ab9515)

Backport 5ab9515 from #10550 --- **What this PR does / why we need it**: Even though the[ Loki HTTP API docs for the push endpoint][1] state that the stream label values should be strings, we previously didn't enforce this requirement. With #9694, we started enforcing this requirement, and that broke some users. In this PR we are reverting this type of assertion and adding a bunch of tests to avoid the regression in the future. [1]: https://grafana.com/docs/loki/latest/reference/api/#push-log-entries-to-loki Co-authored-by: Salva Corts <[email protected]>

**What this PR does / why we need it**: Even though the[ Loki HTTP API docs for the push endpoint][1] state that the stream label values should be strings, we previously didn't enforce this requirement. With grafana#9694, we started enforcing this requirement, and that broke some users. In this PR we are reverting this type of assertion and adding a bunch of tests to avoid the regression in the future. [1]: https://grafana.com/docs/loki/latest/reference/api/#push-log-entries-to-loki

Add metadata to push payload

b799556

pull-request-size bot added the size/M label Jun 13, 2023

salvacorts added 3 commits June 13, 2023 09:52

Add metadata to tests

b0810c8

update vendor push pkg

b7ae131

WIP fix tests

9bec73a

pull-request-size bot added size/L and removed size/M labels Jun 13, 2023

Fix more tests

d876c21

salvacorts commented Jun 13, 2023

View reviewed changes

pkg/logql/log/fmt_test.go Outdated Show resolved Hide resolved

salvacorts added 3 commits June 13, 2023 11:49

Revent some unwanted changes

9c6fdef

Remove more unwanted changes

b4ed160

Update vendor files

8c6f22e

This was referenced Jun 13, 2023

Add non-indexed metadata to chunks #9700

Merged

Metadata to labels result and filtering support #9702

Merged

sandeepsukhani added 6 commits June 16, 2023 11:33

use json format for metadata labels in v1 push api

c854d06

rename metadataLabels to labels

95ba8a2

Merge branch 'main' into salvacorts/metadata-push-payload

7db3e54

lint and some test fixes

ea8f3f6

lint

73cf5dc

lint lint lin...

7b77c69

salvacorts commented Jun 16, 2023

View reviewed changes

integration/client/client.go Outdated Show resolved Hide resolved

sandeepsukhani added 2 commits June 19, 2023 16:09

use map object instead of string representation for log labels in jso…

44b0932

…n push request

revert some accidental chagnes

617e864

owen-d reviewed Jun 22, 2023

View reviewed changes

vlad-diachenko reviewed Jun 23, 2023

View reviewed changes

pkg/push/push.proto Outdated Show resolved Hide resolved

integration/client/client.go Outdated Show resolved Hide resolved

vendor/github.com/grafana/loki/pkg/push/push.proto Outdated Show resolved Hide resolved

owen-d reviewed Jun 23, 2023

View reviewed changes

salvacorts mentioned this pull request Jun 27, 2023

Include entry metadata bytes in metrics tracking ingestion stats #9808

Merged

7 tasks

salvacorts added 3 commits June 29, 2023 10:43

Make labels in loghttp.Entry of type LabelSet

f43dfab

Apply CR feedback

384d556

Avoid allocations for empty labels

4088112

pull-request-size bot added the size/XL label Jul 12, 2023

sandeepsukhani reviewed Jul 12, 2023

View reviewed changes

vendor/github.com/grafana/loki/pkg/push/types.go Outdated Show resolved Hide resolved

vlad-diachenko approved these changes Jul 12, 2023

View reviewed changes

salvacorts and others added 3 commits July 12, 2023 13:10

Update vendor/github.com/grafana/loki/pkg/push/types.go

eaf235c

Co-authored-by: Sandeep Sukhani <[email protected]>

Fix tests

7438809

Don't use yolostring when deserializing push request

f6e6918

sandeepsukhani reviewed Jul 12, 2023

View reviewed changes

pkg/loghttp/query.go Outdated Show resolved Hide resolved

owen-d approved these changes Jul 12, 2023

View reviewed changes

Remove yolostring

f8c33aa

salvacorts merged commit aae13c3 into main Jul 13, 2023

salvacorts deleted the salvacorts/metadata-push-payload branch July 13, 2023 07:13

salvacorts mentioned this pull request Jul 14, 2023

Remove Prometheus dependency for push pkg #9937

Merged

grafanabot mentioned this pull request Jul 17, 2023

[k159] Remove Prometheus dependency for push pkg #9956

Merged

salvacorts mentioned this pull request Sep 12, 2023

Fix regression when parsing numbers in Push request #10550

Merged

grafanabot mentioned this pull request Sep 12, 2023

[k167] Fix regression when parsing numbers in Push request #10551

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add metadata to push payload #9694

Add metadata to push payload #9694

salvacorts commented Jun 13, 2023 •

edited by sandeepsukhani

Loading

owen-d left a comment •

edited

Loading

owen-d Jun 22, 2023

salvacorts Jun 29, 2023

owen-d Jun 22, 2023

salvacorts Jun 29, 2023 •

edited

Loading

sandeepsukhani Jul 5, 2023

owen-d Jun 22, 2023

salvacorts Jun 29, 2023 •

edited

Loading

salvacorts Jun 29, 2023

vlad-diachenko Jun 29, 2023

owen-d Jul 6, 2023 •

edited

Loading

vlad-diachenko Jul 11, 2023

salvacorts Jul 11, 2023

vlad-diachenko left a comment

owen-d Jun 23, 2023

salvacorts Jun 29, 2023

salvacorts Jun 29, 2023

owen-d Jun 23, 2023

salvacorts Jun 29, 2023

owen-d Jun 23, 2023 •

edited

Loading

salvacorts Jun 29, 2023

vlad-diachenko left a comment

vlad-diachenko Jul 12, 2023

sandeepsukhani Jul 12, 2023

salvacorts Jul 12, 2023

vlad-diachenko Jul 12, 2023

salvacorts Jul 12, 2023

salvacorts commented Jul 12, 2023

owen-d left a comment

	entries := make([]logproto.Entry, len(s.Entries), len(s.Entries))
	for i, e := range s.Entries {
	entries[i] = e.ToProto()
	}

Add metadata to push payload #9694

Add metadata to push payload #9694

Conversation

salvacorts commented Jun 13, 2023 • edited by sandeepsukhani Loading

owen-d left a comment • edited Loading

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

salvacorts Jun 29, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

salvacorts Jun 29, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

owen-d Jul 6, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vlad-diachenko left a 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

owen-d Jun 23, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

vlad-diachenko left a 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

salvacorts commented Jul 12, 2023

owen-d left a comment

Choose a reason for hiding this comment

salvacorts commented Jun 13, 2023 •

edited by sandeepsukhani

Loading

owen-d left a comment •

edited

Loading

salvacorts Jun 29, 2023 •

edited

Loading

salvacorts Jun 29, 2023 •

edited

Loading

owen-d Jul 6, 2023 •

edited

Loading

owen-d Jun 23, 2023 •

edited

Loading