OpenSearch Problems & Issues
Real user-reported problems, complaints, and limitations for OpenSearch. Based on analysis of 13 signals.
Overall score: 63/100
Issues by Category
JVM memory pressure and OutOfMemory crashes
AWS documents that heap above about 80% is risky, and at 92% for 30 minutes the managed service blocks writes. Too many shards, fielddata on text fields and heavy aggregations are the usual triggers.
Too many shards: shard limit exceeded
Default limit is 1,000 shards per node; creating an index fails past it. Teams that create daily indices with default shard counts hit this and pay in heap overhead long before.
429 Too Many Requests and rejected bulk indexing
Thread pool queues fill, or on Amazon's service payloads that would exceed heap are throttled. Shows up as 429 errors on bulk indexing and search, often on only some nodes when shards are skewed.
Slow indexing and merge pressure without tuning
Third-party guides warn that default merge and thread pool settings can limit indexing throughput and heap under high concurrency, requiring refresh interval and bulk size tuning.
Upgrades and Elasticsearch migrations are not drop-in
Snapshots from Elasticsearch work smoothly only up to roughly 7.10.x; later versions have diverged, ILM and ISM differ, and in-place managed upgrades cannot be undone without support help.
OpenSearch Dashboards 'server is not ready yet'
One of the most-posted forum errors: Dashboards cannot reach or authenticate to the cluster (wrong hosts or https, blocked port 9200, red cluster, or a stuck saved-object migration).
Managed service limits: no SSH, no opensearch.yml
On Amazon OpenSearch Service you cannot SSH to nodes or edit config files; only some settings are exposed, and limits like search.max_buckets need a support case.
OpenSearch cluster status red or yellow
Red means a primary shard is unassigned; yellow means a replica is. AWS docs list failed nodes and sustained heavy load as the most common causes, and single-node clusters are always yellow. A frequent first fear for new operators.
Disk watermark makes indices read-only
When disk fills past the flood-stage watermark indices get a read-only-allow-delete block, writes fail with ClusterBlockException, and users report having to clear the block manually after freeing space.
Elastic clients and Beats refuse to connect
Default Logstash and Beats distributions include license checks and fail against OpenSearch. You must use Apache 2.0 builds or OpenSearch-compatible clients, a surprise for teams copying Elastic tutorials.
Custom plugins break on every version
Plugins must match the exact OpenSearch version, so each upgrade needs a rebuild; installation fails with a plugin validation error on mismatch.
Docker/quickstart fails: admin password and vm.max_map_count
Since 2.12 the demo security setup refuses to start without a strong OPENSEARCH_INITIAL_ADMIN_PASSWORD, and hosts with vm.max_map_count below 262144 fail bootstrap checks. Common forum threads for first-time installs.
Amazon OpenSearch Service gets expensive
AWS pricing example: three r6g.xlarge data nodes plus three dedicated manager nodes already cost over $1,000 a month before storage. Extended Support fees, data transfer and Serverless OCUs add on top.
Regret Patterns
Common reasons users regret choosing OpenSearch.
Daily indices with default shard counts
Teams pile up thousands of tiny shards, then hit the shard limit and heap pressure.
Picking managed to avoid ops, then paying for it
Multi-AZ domains with dedicated managers and storage tiers cost four figures monthly, and you still tune shards.
Treating it as drop-in Elasticsearch
Clients, Beats and ILM policies need changes.
Running a single node in production
Always yellow, no failover, and any crash risks data.
Ignoring disk until writes block
Flood-stage watermark halts writes and requires manual recovery.
Using OpenSearch for tiny site search
Operating a cluster for a few thousand documents is more work than lighter engines.
When OpenSearch Breaks Down
Specific scenarios where OpenSearch fails or underperforms.
Disk usage passes the flood-stage watermark
Indices go read-only and writes return ClusterBlockException until space is freed and the block cleared.
Heap stays above 85-90%
Long GC pauses, then 429s, write blocks on managed domains, and finally OutOfMemoryError.
Thousands of small shards
Cluster state and heap overhead balloon; new index creation fails at the shard limit.
Node or master loss without enough replicas
Primary shards go unassigned, status turns red, and managed domains can enter read-only after quorum loss.
Major version upgrade with custom plugins
Plugins must match the exact version and need rebuilding; in-place managed upgrades are hard to roll back.
Bulk ingest spike on skewed shards
Some nodes reject requests with 429 while others sit idle.