Additional storage backends

[yurishkuro]

Opening this issue to keep track of other related issues.

• Badger KV (memory/disk) as build-in option Sorted key/value store (badger) backed storage plugin #760
• ScyllaDB Why not use ScyllaDB? #197 (available)
• InfluxDB via grpc-plugin (for v1 and v2 and iOx) InfluxDB as trace storage backend #272 (available)
• Couchbase via grpc-plugin Couchbase as storage backend #1575 (available)
• OpenSearch Official OpenSearch storage option #3044
• ClickHouse ClickHouse as a storage backend #1438
• PostgreSQL PostgreSQL gRPC storage plugin #1895
• Alibaba Cloud Log Service Support Alibaba Cloud Log Service as storage backend  #3172
• RediSearch https://github.com/nicolastakashi/jaeger-redisearch
• Amazon S3 Google Cloud Storage(GCS) and Microsoft Azure Blob Storage support Amazon S3 as a storage backend #2633
• ConsmosDB Comsos DB as storage backend #1667
• MySQL Support mysql as backend ? #1944
• Humio Humio as storage backend #3005
• Quickwit Quickwit as a storage backend #4362
• YugabyteDB Jaeger YugabyteDB(YCQL) support #4220
• Azure Data Explorer (Kusto) [Feature]: Support Azure Data Explorer as storage #4640
• MongoDB [Feature]: Support MongoDB as a storage backend #4712
• ❌ Amazon DynamoDB DynamoDB support #421
• ❌ BigTable Bigtable Support #1208
• ❌ Elasticsearch APM data format Compatibility with Elasticsearch APM UI #1365
• ❌ Netflix Dynomite Support for Netflix Dynomite as storage backend #331
• deprecated PostgreSQL with Promscale

Relevant issue: plugin support #422 (done).

[nbettiol]

Did you remove the flags for elasticsearch in jaeger-collector? Because I'm doing a test using the image docker, which version is:

{"gitCommit":"dbd5db721fc59431b1e64874cc7d6265d89ec917","GitVersion":"v1.1.0","BuildDate":"2018-01-08T21:56:21Z"}

and I cannot see the elasticsearch flags.

[black-adder]

It looks like you're using latest instead of 1.1. We recently moved around some of the flags so that we can support plugins better #625. Using latest, you have to instead use env variable SPAN_STORAGE=elasticsearch to use the elasticsearch flags. I'd recommend that you use 1.1 since this change will be apart of 1.2 and will be documented at that time.

[nbettiol]

Thanks for the reply, yes I was using the latest version. I will use the 1.1

[fzakaria]

I would love to see a SQL option (whatever ANSI SQL that will be least vendor lock-in).
Setting up Cassandra / ElasticSearch might be too ambitious for projects that want distributed tracing but honestly don't have the TPS to warrant a distributed datastore.

[ringerc]

Since I work with PostgreSQL, I sure wouldn't complain. But honestly I'm not sure a SQL db is an optimal store for largely free-form metrics of this nature. PostgreSQL at least offers the jsonb type for indexable free-form data. If you're trying to do this in a vendor neutral way you'll land up with your own json blobs, or doing EAV, and both of those are terrible. ANSI SQL is a poor fit for variable-structured or key/value form data and you'll need some vendor extensions to get usable performance.

But you inevitably land up with someone putting an ORM on top to "abstract" the DB. Then the ORM performs terribly, gobbles memory and everyone says "the SQL backend is slow, use instead".

[pavolloffay]

Related issue to this one is #551. Upvote if you are interested in it.

[SwarnimRaj]

New related issue-
Files - #894

[wy100101]

We are looking at using BigQuery as a storage layer. Presumably this could work with a SQL storage option. SQL can be a generic way to deal with columnar data stores in a generic way. I would complain about a BigQuery specific solution, but I think there is a place for generic SQL interface beyond RDBs.

[yurishkuro]

I assume that even if some database can be treated as SQL and accessed via standard database/sql API, we still need to statically import the actual driver. Granted, this may be less maintenance than a dedicated SpanStorage implementation. However, now that the protobuf model has been merged, nothing is blocking us from moving on the storage plugin dev, eg using something like harshicorp grpc plugin framework.

[isaachier]

Our model is sufficiently simple to warrant looking into using an ORM to support a large number of backends. I'll take a look at what's available. Reread above and understand what @yurishkuro means.

[bruth]

Giving my two cents.. an ANSI SQL could work for small workloads, so may be useful for lower-throughput applications that still want to benefit from this tool.

I will also throw out there that Timescale (a Postgres extension) may be a good fit for the required high write throughput.

[mcarbonneaux]

Clickhouse are SQL high performance storage very efficient for log and trace storage and whold be perfect storage alternative to cassandra original one... they are a true column db... distributed...compressed...

they are near to the CQL (sql like query language)... they use an SQL like language to...

https://clickhouse.yandex/

[chvck]

I just thought that I'd drop something here to say that there is also support for using Couchbase as a storage backend (via the grpc plugin), currently at https://github.com/chvck/couchbase-jaeger-storage-plugin. Will likely move to the couchbase-labs organisation in time.

[omerlh]

Has someone started to work on Azure CosmosDB integration? It has support for Cassandra API, but I couldn't manage to make it work...

[rleiwang]

I just created an issue proposing Chronowave as storage backend. #2534

[DjinNO]

What about ClickHouse? Clickhouse is very cool

[muhammadn]

This is the only issue that is not closed. #2633 is closed so i am announcing my work to support s3 for jaeger as a plugin. We needed a cheap way to store data and elasticsearch is expensive for us (whether managed - price or self-managed - on the operational perspective)

I took the work done by the amazing team at Grafana for their loki code and use that to write an s3 plugin. It's still a WIP and it stores indexes locally and syncs chunks of data on s3. (I am seeing indexes being synced as well).

Data stored is in boltdb format and shipped to s3 via cortex. (i am leveraging on loki's codebase i am not repeating doing other people's work). Cortex is also used to query the data from s3 using PromQL (which i will implement on the plugin's reader)

a side effect of this work is support for Google's GCP, Amazon DynamoDB and Google's BigTable as well.

More info here: https://grafana.com/docs/loki/latest/operations/storage/boltdb-shipper/

jaeger-s3 code: https://github.com/muhammadn/jaeger-s3/tree/develop

Also you will need my fork of jaeger, which is a small change for the file flag at https://github.com/muhammadn/jaeger-s3/tree/develop

Note: It's working on both postgresql and now it can upload to s3. so two things in parallel for me to debug in postgresql to better understand the data. But the postgresql code will be removed once i get both writer and reader working.

Contributions are welcome!

[jkowall]

Interesting idea, good luck with the work @muhammadn. Remember, you'll have to figure out how to support querying for the data. My understanding of loki is that you can't do a free text search easily since the data is not all indexed. How will you support the querying of data with this new storage support?

[muhammadn]

Thanks @jkowall ! Conversations like this is what motivates me. 🙌.

I've managed to pull out the data from s3 through cortex by querying it using promql (basically logql which is based on promql) - You're right that the data indexed only are the time and labels and i can make a search using the labels. Labels are the only thing you can do a search on, though may not as flexible as SQL or ES.

But i am thinking if it's better to go to Grafana's Tempo as i found out about it. But that's a discussion i would make in Tempo's project.

[jkowall]

The creator of Tempo @joe-elliott is a maintainer here at Jaeger and is also an engineer with Grafana Labs. We had this discussion, but right now Tempo cannot support the types of queries the Jaeger UI does today. I know that may change, at which point Tempo would be a good Jaeger backend, but today it cannot do everything necessary.

This is the compromise with something that doesn't do full text indexing versus selective queries or allowing queries based on a specific trace or short timeframe. The cost benefit of using non-indexing technologies are there, but flexibility is not the same.

[muhammadn]

@jkowall I've stabilised the code without any crashes now and I can query/see the data from Jaeger UI. 🎉 So moving forward i will test it out on production (using k8s) and see how it goes from then on. Thanks @jkowall for your answers! 🙌 Also thanks to @yurishkuro for adding plugin support as without it, it would not be possible for me to write the code for s3 support.

I've watched @joe-elliott 's video. Tempo is limited to searching TraceID but with jaeger-s3, searching is limited to labels and there are many label variables that i can use to construct the query.

Discussions on jaeger-s3 plugin will be on my repo, thanks all!

[jkowall]

Cool! Nice job @muhammadn I'm curious how the performance is with a search using S3 in that manner. Are you planning on deploying or using this setup?

[yurishkuro]

@muhammadn I re-opened #2633 and linked your repo at the top. Suggest moving discussions/updates there.

[em135]

@Xitric and I are working in on a storage backend for Humio using the grpc plugin. The repository can currently be found at https://github.com/em135/humio-jaeger-plugin. I have opened an issue for this: #3005

[galan]

If S3 is supported, I would suggest to support GCS as well, which is also an objects-storage for all Google Cloud users. For our use-case that would be tremendous helpful!

[muhammadn]

@galan it does actually, despite the name is jaeger-s3, i had already added support for GCS (and Azure Storage) for quite some time but i have not tested it.

Maybe you can go to https://github.com/muhammadn/jaeger-s3 to try it out.

Related code to GCS:
https://github.com/muhammadn/jaeger-s3/blob/main/config/config.go#L26
https://github.com/muhammadn/jaeger-s3/blob/main/s3store/store.go#L50

All you need is to modify the configuration for jaeger-s3 to this:

https://grafana.com/docs/loki/latest/operations/storage/boltdb-shipper/#example-configuration

which has the GCS config.

I will update the documentation to jaeger-s3 include GCS and Azure as well. (and probably change the project name entirely)

Do tell me if you need help. But i think we can move this discussion to #2633

@galan Update: I have updated the documentation - https://github.com/muhammadn/jaeger-s3/blob/main/README.md

Also just a question from the community, should i rename this project as a more generic name rather than jaeger-s3 since this plugins will support GCS and Azure as well?

[qiansheng91]

@jpkrohling The Alibaba cloud log service has supported the jaeger, and here are the gifs of the plugin. Link: https://github.com/qiansheng91/jaeger-sls#quick-start

[evilrussian]

Can i use other jaeger collector as backend for my jaeger collector?

[pavolloffay]

Jaeger collector cannot send data (e.g. over gRPC) to other jaeger collector. However this capability is supported with OTEL collector.

[nitinsaprumaersk]

@yurishkuro Would highly recommend to add Azure Table Storage as a backend storage option as well.

[arajkumar]

@yurishkuro Could you please add PostgreSQL with Promscale into the list.? Now Promscale is Jaeger storage complaint too :) Thanks.

[nicolastakashi]

@yurishkuro could you please add RediSearch to the list? I've worked on a GRPC Plugin to Store Traces on Redis Search and this is close to the first release, I'm just adding a few performance tests.

https://github.com/nicolastakashi/jaeger-redisearch

[coverthesea]

Zinc https://github.com/zinclabs/zinc

[vemula-anu]

i use jaeger with timescaledb in kubernetes
{"level":"info","ts":1677823818.9793034,"caller":"querysvc/query_service.go:137","msg":"Archive storage not created","reason":"archive storage not supported"}
i use this configurations
spec:
containers:
- env:
- name: GRPC_STORAGE_SERVER
value: promscale:9202
- name: SPAN_STORAGE_TYPE
value: grpc-plugin
image: jaegertracing/jaeger-query:1.30
name: jaeger
ports:
- containerPort: 16685
- containerPort: 16686
restartPolicy: Always
the problem getting can anyone give solution for this

[yurishkuro]

@vemula-anu please do not post support questions to this issue, create a new question in Discussions.

[diondew]

Could you please add Yugabyte to the list? #4354

[jkowall]

Could you please add Yugabyte to the list? #4354

Added, sorry for the delay.

[paulgrav]

The creator of Tempo @joe-elliott is a maintainer here at Jaeger and is also an engineer with Grafana Labs. We had this discussion, but right now Tempo cannot support the types of queries the Jaeger UI does today. I know that may change, at which point Tempo would be a good Jaeger backend, but today it cannot do everything necessary.

I know Tempo today is a lot more capable in terms of its search compared to back in 2021. Is it now a potentially good Jaeger backend or are there still gaps?

[joe-elliott]

Tempo has a broader set of search capabilities (via TraceQL) than Jaeger search. So, generally, Tempo could be used to back Jaeger. There are two gaps I'm aware of.

First, Jaeger search currently returns the entire trace to the frontend which then renders the search results pane you see in the UI. Tempo, however, only returns metadata to the frontend. I don't believe this metadata is enough to render every element of the Jaeger search results. For instance, I don't think you could list the services in the trace.

Second, we currently only retrieve auto complete tags from recent traces. So if you were searching a time range from yesterday the auto complete would still be based on the traces received in the last half hour or so. We are working to address this.

[muhammadn]

@paulgrav

Just what @joe-elliott explained but i want to add that it had been already done. I had completely overhauled jaeger-objectstorage to use tempo as a backend to store traces to AWS S3/GCS/AzureBlob since i believe tempo is already mature to support multiple cloud storage providers. Back in the early days of tempo there wasn't support for GCS and AzureBlob so we used loki as the interface to store trace data.

I've posted in the forums on how it would look like so you can take a look.

The codes are already published (both the tempo fork and jaeger-objectstorage) but the documentation needs more polishing.