Whenever you’re spinning up your Amazon OpenSearch Service area, that you must work out the storage, occasion sorts, and occasion rely; resolve the sharding methods and whether or not to make use of a cluster supervisor; and allow zone consciousness. Usually, we take into account storage as a tenet for figuring out occasion rely, however not different parameters. On this submit, we provide some suggestions primarily based on T-shirt sizing for log analytics workloads.
Log analytics and streaming workload traits
Whenever you use OpenSearch Service on your streaming workloads, you ship knowledge from a number of sources into OpenSearch Service. OpenSearch Service indexes your knowledge in an index that you just outline.
Log knowledge naturally follows a time sequence sample, and due to this fact a time-based indexing technique (every day or weekly indexes) is really useful. For environment friendly administration of log knowledge, you should implement time-based index patterns and set retention durations. You additional outline time slicing and a retention interval for the info to handle its lifecycle in your area.
For illustration, take into account that you’ve got a knowledge supply producing a steady stream of log knowledge, and also you’ve configured a every day rolling index and set a retention interval of three days. Because the logs arrive, OpenSearch Service creates an index per day with names like stream1_2025.05.21, stream1_2025.05.22, and so forth. The prefix stream1_* is what we name an index sample, a naming conference that helps group-related indexes.
The next diagram reveals three main shards for every every day index. These shards are deployed throughout three OpenSearch Service knowledge cases, with one reproduction for every main shard. (For simplicity, the diagram doesn’t present that main and reproduction shards are all the time positioned on totally different cases for fault tolerance.)
When OpenSearch Service processes new log entries, they’re despatched to all related main shards and their replicas within the lively index, which on this instance is barely as we speak’s index because of the every day index configuration.
There are a number of necessary traits of how OpenSearch Service processes your new entries:
- Whole shard rely – Every index sample may have a D * P * (1 + R) complete shards, the place D represents retention in days, P represents main shards, and R is the variety of replicas. These shards are distributed throughout your knowledge nodes.
- Energetic index – Time slicing implies that new log entries are solely written to as we speak’s index.
- Useful resource utilization – When sending a _bulk request with log entries, these are distributed throughout all shards within the lively index. In our instance with three main shards and one reproduction per shard, that’s a complete of six shards processing new knowledge concurrently, requiring 6 vCPUs to effectively deal with a single
_bulkrequest.
Equally, OpenSearch Service distributes queries throughout the shards for the indexes concerned. When you question this index sample throughout all 3 days, you’ll have interaction 9 shards, and want 9 vCPUs to course of the request.
It will get much more difficult if you add in additional knowledge streams and index patterns. For every further knowledge stream or index sample, you deploy shards for every of the every day indexes and use vCPUs to course of requests in proportion to the shards deployed, as proven within the previous diagram. Whenever you make concurrent requests to a couple of index, every shard for all of the indexes concerned should course of these requests.
Cluster capability
Because the variety of index patterns and concurrent requests will increase, you possibly can rapidly overwhelm the cluster’s sources. OpenSearch Service contains inside queues that buffer requests and mitigate this concurrency demand. You possibly can monitor these queues utilizing the _cat/thread_pool API, which reveals queue depths and helps you perceive when your cluster is approaching capability limits.
One other complicating dimension is that the time to course of your updates and queries relies on the contents of the updates and queries. As requests are available in, the queues are filling on the fee you might be sending them. They’re draining at a fee that’s ruled by the accessible vCPUs, the time they tackle every request, and the processing time for that request. You possibly can interleave extra requests if these requests clear in a millisecond than in the event that they clear in a second. You should utilize the _nodes/stats OpenSearch API to observe common load in your CPUs. For extra details about the question phases, check with A question, or There and Again Once more on the OpenSearch weblog.
When you see the queue depths growing, you might be shifting right into a “warning” space, the place the cluster is dealing with load. However if you happen to proceed, you can begin to exceed the accessible queues and should scale so as to add extra CPUs. When you begin to see load growing, which is correlated with queue depth growing, you might be additionally in a “warning” space and may take into account scaling.
Suggestions
For sizing a site, take into account the next steps:
- Decide the storage required – Whole storage = (every day supply knowledge in bytes × 1.45) × (number_of_replicas + 1) × variety of days retained. This accounts for the extra 45% overhead on every day supply knowledge, damaged down as follows:
- 10% for bigger index dimension than supply knowledge.
- 5% for working system overhead (reserved by Linux for system restoration and disk defragmentation safety).
- 20% for OpenSearch reserved area per occasion (phase merges, logs, and inside operations).
- 10% for extra storage buffer (minimizes influence of node failure and Availability Zone outages).
- Outline the shard rely – Approximate variety of main shards = storage dimension required per index / desired shard dimension. Spherical as much as the closest a number of of your knowledge node rely to keep up even distribution. For extra detailed steerage on shard sizing and distribution methods, check with “Amazon OpenSearch Service 101: What number of shards do I would like” For log analytics workloads, take into account the next:
- Really useful shard dimension: 30–50 GB
- Optimum goal: 50 GB per shard
- Calculate CPU necessities – Really useful ratio is 1.25 vCPU:1 Shard for decrease knowledge volumes. Increased ratios are really useful for bigger volumes. Goal utilization is 60% common, 80% most.
- Select the best occasion sort – Contemplate the next primarily based in your nodes:
Let’s take a look at an instance for area sizing. The preliminary necessities are as follows:
- Each day log quantity: 3 TB
- Retention interval: 3 months (90 days)
- Reproduction rely: 1
We make the next occasion calculation.
The next desk recommends cases, quantity of supply knowledge, storage wanted for 7 days of retention, and lively shards primarily based on the previous pointers.
| T-Shirt Measurement | Information (Per Day) | Storage Wanted (with 7 days Retention) | Energetic Shards | Information Nodes | Main Nodes |
| XSmall | 10 GB | 175 GB | 2 @ 50 GB | 3 * r7g.massive. search | 3 * m7g.massive. search |
| Small | 100 GB | 1.75 TB | 6 @ 50 GB | 3 * r7g.xlarge. search | 3 * m7g.massive. search |
| Medium | 500 GB | 8.75 TB | 30 @ 50 GB | 6 * r7g.2xlarge.search | 3 * m7g.massive. search |
| Giant | 1 TB | 17.5 TB | 60 @ 50 GB | 6 * r7g.4xlarge.search | 3 * m7g.massive. search |
| XLarge | 10 TB | 175 TB | 600 @ 50 GB | 30 * i4g.8xlarge | 3 * m7g.2xlarge.search |
| XXL | 80 TB | 1.4 PB | 2400 @ 50 GB | 87 * I4g.16xlarge | 3 * m7g.4xlarge.search |
As with all sizing suggestions, these pointers signify a place to begin and are primarily based on assumptions. Your workload will differ, and so your precise wants will differ from these suggestions. Make certain to deploy, monitor, and modify your configuration as wanted.
For T-shirt sizing the workloads, an extra-small use case encompasses 10 GB or much less of information per day from a single knowledge stream to a single index sample. A small use case falls between 10–100 GB per day of information, a medium use case between 100–500 GB of information, and so forth. Default occasion rely per area is 80 for many of the occasion household. Consult with the “Amazon OpenSearch Service quotas “ for particulars.
Moreover, take into account the next finest practices:
Conclusion
This submit supplied complete pointers for sizing your OpenSearch Service area for log analytic workloads, overlaying a number of important elements. These suggestions function a strong start line, however every workload has distinctive traits. For optimum efficiency, take into account implementing further optimizations like knowledge tiering and storage tiers. Consider cost-saving choices comparable to reserved cases, and scale your deployment primarily based on precise efficiency metrics and queue depths.By following these pointers and actively monitoring your deployment, you possibly can construct a well-performing OpenSearch Service area that meets your log analytics wants whereas sustaining effectivity and cost-effectiveness.
In regards to the authors