I would try to make more broader groups in order to reduce the number of 
shards. I think having 65 small shards creates unnecessary overhead.
Maybe someone with more experience with shared collections can pitch in.

Cheers,

Renato

> On 11 Aug 2026, at 08:42, Harshit Sharma <[email protected]> 
> wrote:
> 
> Hii Renato
> 
> Are you running on fast storage (pd-ssd on GCP)?
> 
> Yes, we are using SSD storage on GCP.
> 
> What is the reason for having 65 shards with only 10-nodes and RF=1, having
>> that many shards is creates overhead.
> 
> Regarding the 65 shards, we have separated the data based on logical
> groupings that align with our data access patterns. This allows us to
> efficiently retrieve data from Solr by using the route parameter, ensuring 
> that
> queries are routed directly to the relevant shard rather than querying
> across all shards.
> 
> RamBuffer of  200MB seems to small for large batches.
> 
> Thanks for the suggestion. We’ll increase the RAM buffer size and monitor
> the metrics to evaluate the impact.
> 
> On Sat, Aug 8, 2026 at 2:41 AM Renato <[email protected]> wrote:
> 
>> Hi Harshit,
>> 
>> Just some thoughts:
>> 
>> Are you running on fast storage (pd-ssd on GCP)?
>> What is the reason for having 65 shards with only 10-nodes and RF=1,
>> having that many shards is creates overhead.
>> RamBuffer of  200MB seems to small for large batches.
>> 
>> Cheers,
>> 
>> Renato
>> 
>>> On 7 Aug 2026, at 12:29, Harshit Sharma <[email protected]>
>> wrote:
>>> 
>>> Hii all
>>> 
>>> We're looking for advice on reducing the time our indexing client waits
>> for
>>> Solr to acknowledge update requests.
>>> 
>>> Our indexing throughput is currently limited primarily by the time taken
>>> for each batched update request to complete. Before changing client-side
>>> concurrency or scaling the application, we'd like to understand whether
>> any
>>> Solr-side optimisations could reduce this acknowledgement latency.
>>> Production Environment
>>> 
>>>  -
>>> 
>>>  SolrCloud 9.6.1
>>>  -
>>> 
>>>  10-node cluster on GCP Compute Engine
>>>  -
>>> 
>>>  Two collections:
>>>  -
>>> 
>>>     ~180M documents (250+ GB, 65 shards)
>>>     -
>>> 
>>>     ~70M documents (180+ GB, 65 shards)
>>>     -
>>> 
>>>  Mostly RF=1
>>>  -
>>> 
>>>  Implicit routing
>>>  -
>>> 
>>>  SolrJ client sending batched update requests (3k-4k documents/request)
>>>  -
>>> 
>>>  Hard commit every *30 minutes* (openSearcher=false)
>>>  -
>>> 
>>>  Soft commit every *15 minutes*
>>>  -
>>> 
>>>  ramBufferSizeMB=*200*
>>>  -
>>> 
>>>  TieredMergePolicy
>>>  -
>>> 
>>>  Default ConcurrentMergeScheduler
>>>  -
>>> 
>>>  StandardDirectoryFactory (MMapDirectory)
>>> 
>>> What we're observing
>>> 
>>> From the client side, a blocking update request typically takes:
>>> 
>>>  -
>>> 
>>>  ~700 ms for batches of 3,000 documents
>>>  -
>>> 
>>>  ~1.5 s for batches of 4,000 documents
>>> 
>>> Occasionally we see several-second outliers, but the median latency
>> remains
>>> consistent even during peak query traffic.
>>> 
>>> We aim to reduce this latency to increase indexing throughput.
>>> Questions
>>> 
>>>  1.
>>> 
>>>  For batched update requests in SolrCloud, what typically contributes
>>>  most to the acknowledgement latency (request distribution, update
>>>  processing, tlog fsync, merges, etc.)?
>>> 
>>>  2.
>>> 
>>>  Which Solr metrics would you recommend checking first to identify where
>>>  this time is being spent?
>>> 
>>>  3.
>>> 
>>>  Are there any Solr configuration settings that commonly have a
>>>  significant impact on update acknowledgement latency (merge policy,
>> merge
>>>  scheduler, indexing buffer, commit settings, update log configuration,
>>>  etc.)?
>>> 
>>>  4.
>>> 
>>>  Has anyone run into similar indexing latency at this scale (60+ shards,
>>>  ~200M documents), and if so, what changes provided the biggest
>> improvement?
>>> 
>>> We'd appreciate any suggestions or best practices to reduce update
>> request
>>> latency and improve indexing throughput. Please let us know if any
>>> additional configuration details or metrics would be helpful.
>>> 
>>> 
>>> Thanks & Regards
>>> 
>>> Harshit Sharma
>> 
>> 
> 
> -- 
> 
> Thanks & Regards
> 
> Harshit Sharma
> 
> Sr. Engineer |  IM-Search

Reply via email to