Performance Baseline · April 14, 2026

Before Marketplace-Aware

Production APM snapshot — last 24h via Datadog + parser docker logs

CSGO Inventory Query (24h)

Avg 349ms · Peak 656ms · Off-peak 158ms

Endpoint Overview

GET /api/v2/inventory/data · 28,350 requests · 0% error rate

250ms

Avg Latency

17s

Max Latency

28.4k

Requests/24h

Latency by Game

CSGO dominates — largest index, HaloSkins merged

CSGO (730)

349ms

TF2 (440)

135ms

RUST (252490)

95ms

Steam (753)

72ms

DOTA2 (570)

4ms

CSGO 730 — Peak vs Off-Peak

Peak: 15:00–21:00 UTC (5–11 PM CET) · Off-peak: 02:00–08:00 UTC

Peak Hours

656ms

3,926 requests · 15:00–21:00 UTC

Off-Peak Hours

158ms

3,641 requests · 02:00–08:00 UTC

4.2x latency increase during peak — EU evening traffic

Hourly Heatmap

Worst: 17:00 UTC (851ms) and 19:00 UTC (826ms)

13:00

443ms

14:00

408ms

15:00

362ms

16:00

396ms

17:00

851ms

18:00

524ms

19:00

826ms

20:00

342ms

21:00

354ms

22:00

301ms

05:00

57ms

06:00

49ms

CSGO 730 — By Query Type

Group Browse

290ms

Default listing · 3,754 req

Unstacking

371ms

Single group · 10,544 req

With Filters

463ms

type/rarity/etc · 515 req

Peak vs Off-Peak by Type

Query Type	Off-Peak	Peak	Multiplier
Group Browse	137ms	642ms	4.7x
Unstacking	168ms	660ms	3.9x

Key Observation

Both query types degrade equally under load. The bottleneck is the monolithic inventory_internal_730 index — all queries compete for the same shard resources during peak. Marketplace-aware split targets this directly.

OpenSearch Span Breakdown

CSGO _internal_ indices are 2–4x slower than non-CSGO

Index Pattern	Avg Latency	Calls/24h	Purpose
`inventory_internal_*`	129ms	39,504	CSGO item replacement + count
`group_internal_*`	76ms	8,149	CSGO group search
`inventory_*` (non-CSGO)	40ms	16,827	TF2/RUST/Steam items
`stickers_internal`	33ms	3,374	Sticker search
`group_*` (non-CSGO)	19ms	2,728	TF2/RUST/Steam groups

Post-deploy: inventory_internal_730 splits into inventory_730 + inventory_730_haloskins

Parser Indexing Cycles

Source: docker logs · 2 PM2 workers (4 PIDs), every 10 min for 730

Game	Items	Groups	Fetch	Parse	Bulk Write	Total
CSGO A	~328k	~15k	14s	14s	~95s	~135s
CSGO B	~353k	~18k	15s	14s	~103s	~145s
RUST	83k	2.2k	0.9s	1.4s	~13s	~17s
TF2	107k	5k	1.3s	2.0s	~21s	~25s
Steam	399k	0.9k	1.3s	6.8s	~72s	~81s

CSGO 730 Bulk Write Variance

Range across 20+ cycles · ~328k items per write (Worker A)

Best

69s

Avg

95s

Worst

135s

Throughput: ~328k docs in ~95s = ~3,450 docs/s

Worst times at :27/:37 when RUST/TF2 overlap with 730 bulk write — shard contention

Known Issue in Current Prod

mustNot camelCase silently ignored by OpenSearch

queryCountItemsOnGroup sends "mustNot" instead of "must_not". Reserved-asset filtering is not applied to item counts.

Fixed in marketplace-aware branch (commit 1ca70ac8)

What to Compare Post-Deploy

Metric	Current Baseline	Target
CSGO 730 avg latency	349ms	Same or lower
CSGO peak latency	656ms	Lower (split indices)
OS inventory_internal avg	129ms	Replaced by 2 smaller indices
OS group_internal avg	76ms	Replaced by 2 smaller indices
Non-CSGO latency	40–95ms	Unchanged
Parser 730 cycle time	~135–145s	Own lane only; HaloSkins separate
Error rate	0%	Stay at 0%

Generated April 14, 2026 · Datadog APM + docker logs · tradeit.gg