How much does it cost to run a search backend?

Search clusters are priced on nodes, not on queries, which means the cost is set the day you choose a topology and barely moves afterwards. Five million queries a month or fifty million, the three data nodes cost the same. That makes search one of the easiest workloads to cost accurately and one of the easiest to overprovision.

The workload we are pricing

Assume a product catalog and document search in us-east-1 with 20 million documents averaging 2 KB, so about 40 GB of source data. Indexed with analysis, stored fields, and one replica, that is roughly 120 GB of index across the cluster. Queries run at 5 million a month, peaking at about 25 per second. Documents update at 500,000 a day, arriving through a queue and indexed in batches. There is an API tier in front of the cluster that handles query parsing, permission filtering, and result shaping.

The monthly breakdown

Indexing is almost free, serving is not

Half a million document updates a day, batched fifty to an invocation, costs 2.56 dollars of compute plus 6 dollars of queue charges. Meanwhile the cluster that holds the result costs 675 dollars. The asymmetry is the defining feature of search economics: you pay to keep an index resident in memory and available, not to build it.

This has a practical consequence. Reindexing, which teams often avoid because it feels expensive, costs almost nothing in compute. A full rebuild of 20 million documents is maybe 400,000 batched writes, a few dollars. The constraint on reindexing is cluster headroom during the rebuild, not money. Sizing the cluster with enough spare capacity to run a blue-green reindex is a far better use of node budget than avoiding reindexing altogether.

Managed against self-managed

The managed premium is about 60 percent, and most of it is the dedicated master nodes, which the managed service strongly recommends and which a self-managed three-node cluster typically co-locates. Whether 380 dollars a month is a fair price for not doing version upgrades, not building snapshot automation, and not being paged when a node fails depends on how many clusters you run and how much search matters to the product. The general shape of the decision is covered inmanaged versus self-hosted search cost.

Right-sizing the cluster

Three r6g.large nodes give 48 GB of total RAM for a 120 GB index. The rule of thumb is that you want the frequently-accessed portion of the index in the filesystem cache, which for a catalog with a hot subset is usually fine. If queries are slow and CPU is low, the cluster is memory-starved and needs bigger nodes, not more of them.

The last two rows cost the same and solve different problems. Scaling up helps when the index does not fit in memory; scaling out helps when queries queue behind each other. Diagnosing which one you have before buying either is worth more than the price difference. Further levers are inOpenSearch cost optimization.

Unit economics

At 946 dollars for 5 million queries, search costs 0.19 dollars per thousand queries. Because the cluster cost is fixed, that unit number falls fast with volume: at 25 million queries a month on the same cluster it is 0.038 dollars per thousand. The implication is that adding search surfaces to a product is nearly free once the cluster exists, while the first search feature carries the whole 675 dollars. Price the topology from Terraform against the resource catalog before committing, because node count and instance family are the only decisions that really matter.

FAQ

What to do next

Choose your cluster topology with the price in front of you. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.