TFS 2017 工作项搜索索引异常问题求助

Let's tackle this tricky search indexing inconsistency you're facing—since you've already tried reconfiguring the search component, we'll dive into more targeted checks and fixes tailored to your TFS 2017 Update 3 + Elasticsearch 2.4.1 setup.

1. Verify Elasticsearch Index Health & Corruption #

First, rule out underlying issues with your Elasticsearch indices, which are often the root of inconsistent search results:

• Log into your Elasticsearch server and run this cluster health check command:

  curl -XGET 'http://<YOUR_ES_SERVER>:9200/_cluster/health?pretty'
  

  You want the status field to show green—if it's yellow or red, there's a problem with shard allocation or cluster stability.
• Check the specific TFS work item indices (they usually start with tfs_workitem_*):

  curl -XGET 'http://<YOUR_ES_SERVER>:9200/_cat/indices/tfs_workitem*?v'
  

  Look for indices with a status other than open, or docs.count values that don't align with your actual work item volume.
• If you suspect minor corruption, run this cleanup command to expunge deleted documents and repair index inconsistencies:

  curl -XPOST 'http://<YOUR_ES_SERVER>:9200/tfs_workitem_*/_forcemerge?only_expunge_deletes=true'
  

2. Audit TFS Search Indexer Jobs & Logs #

TFS relies on background jobs to sync work items to Elasticsearch—stuck or failing jobs can cause partial indexing:

• Open the TFS Administration Console, navigate to Application Tier > Job Agent, and look for jobs labeled Work Item Search Indexing. Check for failed runs, long wait times, or stuck jobs.
• Enable verbose logging to capture detailed errors:
  1. In the TFS Admin Console, go to Search > Search Administration.
  2. Click Change Logging Level and set it to Verbose.
  3. Reproduce your search issues, then check the logs at <TFS_INSTALL_DIR>\Application Tier\Logs for entries tagged with Search or Elasticsearch. Look for errors like failed document syncs, timeout messages, or invalid field mapping warnings.

3. Validate Text Analysis & Field Mappings #

The weird discrepancy between single-term and phrase search results points to potential issues with how Elasticsearch processes your work item text:

• Check the mapping for text fields (like System.Description or System.History) that contain your search terms:

  curl -XGET 'http://<YOUR_ES_SERVER>:9200/tfs_workitem_*/_mapping?pretty'
  

  Ensure these fields are mapped as text (not keyword) and use an analyzer that supports phrase searches (like the default standard analyzer). Keyword fields won't break text into terms, so phrase searches will fail.
• Test how Elasticsearch analyzes your problematic search phrase directly:

  curl -XPOST 'http://<YOUR_ES_SERVER>:9200/_analyze?pretty' -d '{
    "analyzer": "standard",
    "text": "One or more hosts are not enabled by filter"
  }'
  

  Compare the generated tokens to what you expect—if terms are being dropped or split unexpectedly, this could explain missing results.

4. Perform a Full, Clean Index Rebuild #

Since your initial reconfiguration didn't work, let's do a complete rebuild that wipes all cached and indexed data:

1. Stop the TFS Job Agent service on your application tier server.
2. On the Elasticsearch server, delete all TFS-related indices:

   curl -XDELETE 'http://<YOUR_ES_SERVER>:9200/tfs_*'
   

3. Delete the TFS search cache folder (default path: <TFS_DATA_DIR>\Search\Cache).
4. Start the TFS Job Agent service again.
5. In the TFS Admin Console, go to Search > Search Administration, then click Start Full Index for work item search.
6. Wait for the full index to complete—this can take hours depending on how many work items you have, so don't interrupt it.

5. Check for Compatibility Patches #

TFS 2017 Update 3 and Elasticsearch 2.4.1 have known compatibility quirks that were fixed in later patches:

• Install all available patches for TFS 2017 Update 3 from the Microsoft Update Catalog—look for updates released after 2018 (since Update 3 launched in late 2017).
• Upgrade Elasticsearch 2.4.1 to the final patch in the 2.4 branch (2.4.6)—older versions had bugs related to phrase search accuracy and index stability.

内容的提问来源于stack exchange，提问作者Alexey Soloviev