possible memory leak in structured metadata

[shimasima2323]

Describe the bug
I'm trying to push log records into loki via push JSON API, if log records contains structured metadata, loki memory usage increases over time with no limit, i played with some configs like using s3 backend, or changing cache settings with no success.

To Reproduce

1. Started Loki with this config:

auth_enabled: false

server:
  http_listen_port: 3100
  grpc_server_max_recv_msg_size: 104857600
  grpc_server_max_send_msg_size: 104857600

common:
  path_prefix: /loki
  replication_factor: 1
  ring:
    kvstore:
      store: inmemory

schema_config:
  configs:
    - from: 2020-10-24
      store: tsdb
      object_store: s3
      schema: v13
      index:
        prefix: index_
        period: 24h

storage_config:
  tsdb_shipper:
    active_index_directory: /loki/index
    cache_location: /loki/index_cache
    resync_interval: 5s
    shared_store: s3
  aws:
   s3: s3://admin:supersecret@minio.:9000/loki
   s3forcepathstyle: true

query_scheduler:
  max_outstanding_requests_per_tenant: 32768

querier:
  max_concurrent: 16

limits_config:
  ingestion_rate_mb: 100
  ingestion_burst_size_mb: 100
  reject_old_samples: false
  allow_structured_metadata: true

ruler:
  alertmanager_url: http://localhost:9093

analytics:
 reporting_enabled: true

2. Started to send log batches to push api '...'

Expected behavior
the ingestions continues with proper memory consumption.

Environment:

• Infrastructure: docker container on debian host
• Deployment tool: docker compose

[davinkevin]

I have the same problem in loki 2.9.1 using minio as backend, tsdb but nothing related to metadata.

I use almost the default configuration from the chart: https://gitlab.com/davinkevin.fr/home-server/-/blob/3834527c8db3b2e9c864419ea65362f508dbf06d/monitoring/overlays/k8s-server/loki/loki.values.yaml

Note

Tested with loki 2.9.2, same problem. it consumed 20Gb, killing it doesn't solve it…

[DylanGuedes]

do you see any errors in your logs? wondering if it isn't that ingesters are failing to flush data

[davinkevin]

Chicken and egg problem, I had to move back to boltdb database and so I have no logs anymore 😓.

[DylanGuedes]

you can't just "change" your database type like that. you should instead add a new config just for boltdb and set the start period as today

[davinkevin]

you can't just "change" your database type like that. you should instead add a new config just for boltdb and set the start period as today

I know and I did… and the system started to consume all CPU and Memory again 😅

[Godzillas]

same problem with loki 2.9.2
When I set up promtail, the memory of the Loki write component started to soar. In fact, I had already set the scheme version to v13 two days ago.

[elcomtik]

Is there any new info? I want to migrate to schema v13 & use structured metadata. I would also like to migrate to TSDB. If there is a memory leak it would be a problem for me.

[davinkevin]

No progress from me.
With boltdb being deprecated and the new loki v3, I'll give it a new try but my setup hasn't changed.

[JohanLindvall]

We see a similar issue in Loki 3 too. I have attached a heap pprof SVG demonstrating the issue.

[JohanLindvall]

Same as #13123, which is now fixed?