Search Functionality Issues with Bleeve Engine

[amix307]

Description

Hi,

I'm following up on my previous ticket: #30064. I always keep the service updated to the latest version, and the search problem persists even on version 1.22.1. I'm using the inbuilt Bleeve engine.

Here’s another example of the issue:
When searching for the string "services.gradle.org", there is one file in the repository where this string should be found. However, the Exact search method does not find the result, and the Fuzzy search hangs for about 2-3 minutes without finding anything.

To give more context about my instance: I have about 300 organizations and 3000 repositories, but overall the size is small as I don't have heavy files. I know for sure that there are many occurrences of the search string across default branches, likely several hundred. Despite this, the Exact search method returns only about 10 results and does so instantly. The Fuzzy search, however, hangs indefinitely. On version 1.21.11, I even encountered a 500 error and the Gitea service restarted.

Manual re-indexing does not help. I would like to resolve the search issues and have more transparent ways to understand how the code is indexed and to have more flexible control over Bleeve settings.

Thanks in advance for your help. I've attached some screenshots for reference.

Gitea Version

1.22.1

Can you reproduce the bug on the Gitea demo site?

No

Log Gist

No response

Screenshots

Git Version

2.31.1

Operating System

CentOS Stream 9

How are you running Gitea?

Self-Hosted from dl.gitea.org

Database

PostgreSQL

[silverwind]

the Fuzzy search hangs for about 2-3 minutes without finding anything.

I also noticed this search hanging starting with v1.22. If kept running, gitea continously consumes more and more memory up until the point where it exhausts the system resources. It's only for certain search terms, not all of them.

[amix307]

the Fuzzy search hangs for about 2-3 minutes without finding anything.

I also noticed this search hanging starting with v1.22. If kept running, gitea continously consumes more and more memory up until the point where it exhausts the system resources. It's only for certain search terms, not all of them.

Yes, yesterday my instance hangs and service gitea was restarted with OOM on VM 64gig ram)

[MICCustomsSolutions]

Same issue here. Eats all the RAM.
Disabling the index fixes the issue.

[carobme]

I can confirm this issue. Using the fuzzy code search with v1.22.1 (self-hosted) results in Gitea memory usage growing until the process runs out of memory.

Rebuilding the index (by stopping Gitea, rm -rf indexers/* and starting again) doesn't make a difference.

[techknowlogick]

Pinging @6543

[makar112233]

We're also had same issue

[smartEBL]

We are affected by this issue as well. A temporary fix for us is changing

[indexer]
REPO_INDEXER_ENABLED = true

to

[indexer]
REPO_INDEXER_ENABLED = false

Code search on single repository still works then, global search does not (of course). So it would be really nice to see that fixed. Is there anything we could provide for debugging?

[amix307]

@smartEBL now i switched to single-node elasticsearch in container at same vm and it works good, but have same issues with non transparent index mechanism (not searching all)

[nekdan]

We also encountered this issue and disabled the search functionality to ensure normal operation. However, this is a poor solution because we need the search functionality for our work.

[lunny]

Can anyone reproduce it in the development version? I think it's related to bleve versions.

[kemzeb]

Can anyone reproduce it in the development version? I think it's related to bleve versions.

I am not yet familiar with how the code search indexing works nor am I familiar with bleve, but this could be the case after looking into the recent bleve releases.

In Gitea v1.21 we use bleve v2.30.10. Gitea v1.22 and onwards uses bleve v2.4.0.

bleve v2.4.1 adds a fix to a memory leak problem associated to their "vector query path" (more specifically, it was a problem in a indirect dependency called blevesearch/go-faiss).

I am not sure if this "path" is something that our code will eventually execute (or if it is only used by the vector indexing feature introduced in bleve v2.4.0), but I wish to bring this up to those that are maybe more familiar.

[silverwind]

According to go.mod, gitea v1.22 still uses bleve v2.3.10, so exact same version as v1.21:

https://github.com/go-gitea/gitea/blob/release/v1.22/go.mod#L22
https://github.com/go-gitea/gitea/blob/release/v1.21/go.mod#L21

I think the issue must lie in first-party code.

[kemzeb]

After doing the following when firing up my dev Gitea instance:

• Create an empty repo
• Push Gitea's entire commit history to this empty repo (chose Gitea since its has a large history)
• Execute a fuzzy code search with the query "merge.go"

I was able to notice the huge memory cost. Here is my generated pprof graph for reference.

I also checked out snapshot 1262ff6 (as this was before major changes were made to code search fuzzing) and I did not notice a performance impact on memory when observing my heap usage in the admin dashboard.

Don't have time to dig into this further yet, but thought this could be helpful in some way.

[lunny]

Bleve has been upgraded to v2.4.2 via #31762 but I don't whether this issue has been resolved.

[amix307]

Bleve has been upgraded to v2.4.2 via #31762 but I don't whether this issue has been resolved.

i can try to test on nightly 1.22

[silverwind]

@smartEBL now i switched to single-node elasticsearch in container at same vm and it works good, but have same issues with non transparent index mechanism (not searching all)

I switched to elasticsearch repo indexer as well now. It seems to be more efficient regarding disk usage than bleve, likely because of missing compression of the bleve index files. I noticed one bug so far in that exact search does not work correctly with elasticsearch, it's always fuzzy.

[smartEBL]

@amix307 @silverwind
How do your setups look like now? How did you change the app.ini to achieve that?

[amix307]

@amix307 @silverwind How do your setups look like now? How did you change the app.ini to achieve that?

[indexer]

; before:

;ISSUE_INDEXER_NAME = gitea_issues
;ISSUE_INDEXER_TYPE = bleve
;MAX_FILE_SIZE = 1048576
;REPO_INDEXER_ENABLED = false
;REPO_INDEXER_EXCLUDE = resources/bin/**
;REPO_INDEXER_INCLUDE =
;REPO_INDEXER_PATH = indexers/repos.bleve
;REPO_INDEXER_TYPE = bleve
;STARTUP_TIMEOUT = 30s

; now:

REPO_INDEXER_ENABLED = true
REPO_INDEXER_REPO_TYPES = sources,forks,mirrors,templates
REPO_INDEXER_TYPE = elasticsearch
REPO_INDEXER_CONN_STR = http://elastic:changeme@localhost:9200
REPO_INDEXER_NAME = gitea_codes
REPO_INDEXER_INCLUDE =
REPO_INDEXER_EXCLUDE =
MAX_FILE_SIZE = 1048576

[smartEBL]

@amix307 @silverwind How do your setups look like now? How did you change the app.ini to achieve that?

[indexer]

; before:

;ISSUE_INDEXER_NAME = gitea_issues
;ISSUE_INDEXER_TYPE = bleve
;MAX_FILE_SIZE = 1048576
;REPO_INDEXER_ENABLED = false
;REPO_INDEXER_EXCLUDE = resources/bin/**
;REPO_INDEXER_INCLUDE =
;REPO_INDEXER_PATH = indexers/repos.bleve
;REPO_INDEXER_TYPE = bleve
;STARTUP_TIMEOUT = 30s

; now:

REPO_INDEXER_ENABLED = true
REPO_INDEXER_REPO_TYPES = sources,forks,mirrors,templates
REPO_INDEXER_TYPE = elasticsearch
REPO_INDEXER_CONN_STR = http://elastic:changeme@localhost:9200
REPO_INDEXER_NAME = gitea_codes
REPO_INDEXER_INCLUDE =
REPO_INDEXER_EXCLUDE =
MAX_FILE_SIZE = 1048576

Cool, thanks for the fast reply. And how complicated is the setup/configuration of the elastic search? You also just spawn it inside some docker container? Or were there any big showstoppers you came across?

[silverwind]

We're getting offtopic, but it's pretty trivial, for example with docker run:

docker run --name="elasticsearch" \
  --volume="/opt/elasticsearch:/usr/share/elasticsearch/data" \
  --volume="/usr/share/zoneinfo:/usr/share/zoneinfo:ro" \
  --volume="/etc/localtime:/etc/localtime:ro" \
  --env="discovery.type=single-node" \
  --env="xpack.security.enabled=false" \
  --env="ingest.geoip.downloader.enabled=false" \
  --env="ELASTIC_USERNAME=elastic" \
  --env="ELASTIC_PASSWORD=changeme" \
  --publish="0.0.0.0:9200:9200" \
  --restart="always" \
  --detach="true" \
  "elasticsearch:7.17.23"

[planbnet]

I just had this problem with a simple search on our local instance running Gitea 1.22.1.
The installation has a few hundred repos and Bleve configured like this:

[indexer]
REPO_INDEXER_ENABLED = true
REPO_INDEXER_PATH = indexers/repos.bleve
MAX_FILE_SIZE = 1048576
REPO_INDEXER_INCLUDE =
REPO_INDEXER_EXCLUDE = client/bower_components/**,bower_components/**,node-modules/**,client/node-modules/**

16 GB of memory are slowly eaten up by the following goroutine (taken from Gitea's monitoring/stacktrace view). Clicking the trash can icon to kill the process does nothing.

GET: /explore/code?q=build.yaml&fuzzy=true

• [...until here, there are always different function calls listed with each reload of the monitoring view, but the following stacktrace items will always be the same..]

• github.com/blevesearch/bleve/v2/search/searcher.(*PhraseSearcher).Next
/go/pkg/mod/github.com/blevesearch/bleve/v2@v2.3.10/search/searcher/search_phrase.go:229

• github.com/blevesearch/bleve/v2/search/collector.(*TopNCollector).Collect
/go/pkg/mod/github.com/blevesearch/bleve/v2@v2.3.10/search/collector/topn.go:228

• github.com/blevesearch/bleve/v2.(*indexImpl).SearchInContext
/go/pkg/mod/github.com/blevesearch/bleve/v2@v2.3.10/index_impl.go:580

• code.gitea.io/gitea/modules/indexer/code/bleve.(*Indexer).Search
/source/modules/indexer/code/bleve/bleve.go:285

• code.gitea.io/gitea/modules/indexer/code.PerformSearch
/source/modules/indexer/code/search.go:138

• code.gitea.io/gitea/routers/web/explore.Code
/source/routers/web/explore/code.go:80

[johanvdw]

Bleve has been upgraded to v2.4.2 via #31762 but I don't whether this issue has been resolved.

This is not solved on gitea 1.22.2

[planbnet]

Just tested it with 1.22.2, still the exact same issue.

To prevent unintended DOS attacks on our server, I've disabled the fuzzy search with a small nginx hack:

location /explore/code {
  if ($arg_fuzzy = 'true') {
    rewrite ^/explore/code /explore/code?q=$arg_q&fuzzy=false last;
  }
  proxy_pass http://localhost:3000;
}

If there's a cleaner workaround to keep the indexer, but disable fuzzy search, please let me know.