VSS Warehouse — Helm Deploy
When to Use
- Deploy, upgrade, or size the VSS warehouse blueprint (2D / 3D / MV3DT) on Kubernetes via Helm
- Compute a GPU-aware
NUM_STREAMS cap for a warehouse Helm install so it matches what the perception pipeline can sustain
Do not use this skill for:
- Docker Compose warehouse deployment — use
vss-build-vision-ai's
references/warehouse.md; it owns the
HARDWARE_PROFILE → GPU mapping table and the blueprint_config.yml stream-cap semantics
this skill reuses.
- Non-warehouse Helm profiles (
base, search, lvs, alerts) — those don't have a
bp-configurator GPU-aware stream cap; deploy them per their own chart READMEs.
- Runtime operations (adding cameras, querying behavior analytics) — use
vss-manage-alerts /
vss-query-analytics against the running deployment.
Why this exists
Docker Compose's warehouse deploy caps NUM_STREAMS per GPU automatically: the configurator reads
deploy/docker/industry-profiles/warehouse-operations/blueprint-configurator/blueprint_config.yml's
max_streams_supported table for the detected HARDWARE_PROFILE and mode, and clamps
final_stream_count = min(NUM_STREAMS, max_streams_supported).
The Helm charts (deploy/helm/industry-profiles/warehouse-operations/warehouse-{2d,3d,mv3dt}-app)
do not do this — their bp-configurator.env ships a fixed NUM_STREAMS and never sets
HARDWARE_PROFILE at all (ENABLE_PROFILE_CONFIGURATOR=false). A user who asks for more streams
than the GPU can sustain gets no protection. This skill closes that gap by computing the same cap
Compose would apply and writing it into a Helm values-override file before install.
Available Scripts
| Script |
Purpose |
Arguments |
../../../deploy/helm/industry-profiles/warehouse-operations/scripts/compute_stream_cap.py |
Detect GPU (or take an explicit HARDWARE_PROFILE), read max_streams_supported from blueprint_config.yml, cap the requested stream count, and write a bp-configurator.env-patched values-override YAML. Pass any values file(s) your install already uses via -f so custom bp-configurator.env entries in them aren't dropped. |
--mode {2d,3d,mv3dt} --num-streams N [--hardware-profile P] [--gpu-index I] [-f VALUES]... [-o FILE] |
This script has no skill/agent dependency — a user who doesn't want to use this skill can run it
directly (python3 compute_stream_cap.py --mode 2d --num-streams 8) and pass the generated file to
helm upgrade/install -f themselves.
Instructions
Precheck the cluster and required inputs before touching Helm — don't assume a fresh
cluster already has these. Run each check and report pass/fail back to the user:
kubectl cluster-info # cluster reachable
kubectl get nodes # all nodes Ready
kubectl get storageclass # a StorageClass exists
kubectl get nodes -o jsonpath='{.items[*].status.allocatable.nvidia\.com/gpu}{"\n"}'
# non-empty -> GPU Operator has registered GPUs
helm version --short # Helm 3.x
Also ask whether the user already has an NGC API key — that can't be checked from cluster
state, only asked about.
On any failure, don't just link the user to the README and stop — hand them the actual fix,
copied from the chart README, and offer to run it for them:
- No
StorageClass → relay the local-path-provisioner install + kubectl patch storageclass
snippet from warehouse-<mode>-app/README.md §Prerequisites (bare-metal option) — or ask
what StorageClass they intend to use if they already have one in mind. Multi-node cluster:
local-path's node affinity can strand vss-vios-nvstreamer's PVCs across different nodes
(didn't match PersistentVolume's node affinity) — relay the same section's
nfs-subdir-external-provisioner snippet instead, and set vios.vstStorage.vstData,
.vstVideo, and .streamerVideos .storageClass to nfs-client via three separate --set
flags (or just global.storageClass) rather than local-path.
- No
nvidia.com/gpu allocatable → relay the NVIDIA GPU Operator install steps from
§Prerequisites (links to the GPU Operator getting-started guide) and the recommended driver
versions listed there.
- Cluster unreachable / nodes not
Ready → this one the user has to fix outside Helm/this
skill entirely; say so plainly rather than suggesting a chart-level fix.
- No NGC API key → point at §Required secrets in the chart README for how to create the pull
secret, don't just say "get an NGC API key."
Only proceed to step 2 once cluster/StorageClass/GPU-Operator/Helm all pass and the user has
confirmed they have an NGC API key — an install started before that will fail partway through
in a way that's harder to debug than catching it here.
Ask ingress vs. NodePort — this determines both what's installed in this step and which
install command gets used in step 6, so resolve it before going further, don't default silently
to one or the other:
- Ingress (needed off-cluster / for a stable hostname) → check whether an ingress
controller is already installed (
kubectl get ingressclass). If not, relay the
haproxy-ingress install snippet from warehouse-<mode>-app/README.md §"Install the ingress
controller" and offer to run it. Note this is a one-time, per-cluster step, not per-app.
- NodePort (simplest for a quick local/single-node deploy, no ingress controller needed) →
tell the user the chart ships
values-nodeport.yaml for this — the install command in step 6
changes to -f values-nodeport.yaml layered under the stream-cap file, and the service URLs
move to <NODE_IP>:<port> instead of <NODE_IP>/<path>. See §"No ingress controller:
NodePort" and §URLs in the chart README for the exact ports.
If the user hasn't said which they want and there's no clear signal (e.g. "just get it running
locally" implies NodePort; "expose it for the team" implies Ingress), ask rather than guessing.
Determine mode and whether to enable Alerts:
- Mode. Use
2d, 3d, or mv3dt if the request already names one. Otherwise ask —
don't guess:
2d — 2D object detection & tracking.
3d — standalone RTVI-CV-3D / multi-camera 3D tracking on calibrated inputs.
mv3dt — Multi-View 3D Tracking warehouse profile. Also needs
rtvi.vss-rtvi-cv.standaloneWarehouse.mv3dt.fusion.maxExpectedSensors set to the effective stream
count in step 6/7 (default 4) — it's BEV fusion's own camera-count setting, separate from
NUM_STREAMS/syncFileCount, and the stream-cap script doesn't touch it.
- Alerts. Not a fourth mode — an optional overlay, off by default, and only available on
2d (warehouse-2d-app is the only chart with vss-alert-bridge/agent/vss-agent-ui as
dependencies; 3d and mv3dt don't have them). If the user is on 3d/mv3dt and asks for
Alerts, say it's not available there instead of trying to enable it. On 2d, ask the user
whether they want it, and explain the tradeoff first rather than enabling or skipping it
for them: without Alerts they get the raw RT-CV detection/tracking stream; with it, detections
also pass through a behavior-analytics stage and a VLM verification step (RT-VLM) before
anything is surfaced as an incident, queryable through the agent/agent UI. That verification
step is the reason to turn it on — it's what keeps every raw detection from becoming a ticket.
If they want it, note the four flags have to be set together
(vss-alert-bridge.enabled, agent.enabled, vss-agent-ui.enabled,
rtvi.vss-rtvi-vlm.enabled — swap the last for an external vlmBaseUrl if not using the
in-cluster VLM) plus Kafka/Elasticsearch/VST endpoint values. Full block:
warehouse-2d-app/README.md §Alerts — layer it in during step 6.
- Stream count. Ask if not given; it sizes the
NUM_STREAMS cap in step 5.
Ask whether the install customizes bp-configurator.env (extra env vars, different
defaults) — don't assume none exist just because the user didn't mention one. If they're
unsure, ask them to check their existing helm upgrade --install command for anything touching
bp-configurator.env, file-based or inline. State the outcome back to them either way:
- Values file (
-f my-values.yaml) → note its path. It gets passed to the script via -f
in the next step and to helm itself in step 7 — the script's output only carries
bp-configurator.env, so anything else in that file (storage class, ingress, alerts flags)
still needs helm to see the original file directly. See
references/streams.md.
- Inline (
--set/--set-json on bp-configurator.env) → the script only reads YAML files,
it can't consume a --set string. Move it into a values file first — see
references/streams.md
for the helm get values -a command (secrets included, handle with care) and why it can't be
trimmed. Then treat it as the values-file case above.
- No customizations → say so explicitly (e.g. "no custom
bp-configurator.env overrides,
so nothing extra is needed here") and proceed without any of the above.
Run the stream-cap script from the repo root:
python3 deploy/helm/industry-profiles/warehouse-operations/scripts/compute_stream_cap.py \
--mode <mode> --num-streams <N> -o values-stream-cap.generated.yaml
- If step 4 found a customizing values file, pass it here too via
-f — otherwise the
generated file (built from chart defaults, layered last) silently drops those customizations.
See
references/streams.md.
- Without
--hardware-profile, it runs nvidia-smi on GPU index 0 and maps the name to a
HARDWARE_PROFILE using the same table as vss-build-vision-ai's warehouse
reference. If detection
fails or the GPU isn't in that table, pass --hardware-profile explicitly. IGX-THOR/
DGX-SPARK edge devices aren't supported by this Helm path.
- No local
nvidia-smi (running helm/kubectl from a bastion, laptop, or CI runner rather
than a GPU node): kubectl exec into a GPU Operator daemonset pod (driver or
device-plugin, e.g. kubectl get pods --all-namespaces -l app=nvidia-driver-daemonset) and
run nvidia-smi --query-gpu=name --format=csv,noheader there instead, then map the name and
pass --hardware-profile.
- It prints the effective (possibly capped) stream count and the
syncFileCount value to keep
in step (see references/streams.md for why).
- It never lowers the request silently without saying so — a cap is always logged to stderr.
Prepare the rest of the values — secrets, storage class, either ingress/externalHost or
the NodePort values file per the choice made in step 2, and — if Alerts was enabled in step 3 —
the four-flag Alerts values block from warehouse-2d-app/README.md §Alerts (Kafka/
Elasticsearch/VST endpoints included). On mv3dt, also add
--set rtvi.vss-rtvi-cv.standaloneWarehouse.mv3dt.fusion.maxExpectedSensors=<effective-streams> (same
value as syncFileCount from step 5). If step 4 found a customizing values file, it goes here
too (-f my-values.yaml) — passing it only to the script in step 5 covers bp-configurator.env
but drops everything else in that file from the install. See
references/streams.md for the full helm upgrade --install command
with the generated file layered in last via -f.
Install/upgrade, chaining the generated file after any other -f/--set overrides so it
wins on bp-configurator.env. The base command is the same either way; only the
ingress-vs-NodePort overrides differ:
helm dependency update deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app
# Ingress:
helm upgrade --install wh deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app \
-n <namespace> --create-namespace \
--set global.vssIngress.enabled=true \
--set global.externalHost=<NODE_IP> \
--set global.storageClass=<STORAGE_CLASS> \
--set vios.vss-vios-nvstreamer.syncFileCount=<effective-streams> \
--set vios.vss-vios-nvstreamer.rtsp.instanceCount=<effective-streams> \
... \
-f values-stream-cap.generated.yaml # last: wins on bp-configurator.env
# NodePort:
helm upgrade --install wh deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app \
-n <namespace> --create-namespace \
-f deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app/values-nodeport.yaml \
--set global.storageClass=<STORAGE_CLASS> \
--set vios.vss-vios-nvstreamer.syncFileCount=<effective-streams> \
--set vios.vss-vios-nvstreamer.rtsp.instanceCount=<effective-streams> \
-f values-stream-cap.generated.yaml # last: wins on bp-configurator.env
... is the remaining secrets/URL overrides from step 6 — see
references/streams.md.
-f values-stream-cap.generated.yaml has to be the last -f in the command — that's what
makes it win on bp-configurator.env (multiple -f files merge in order given, later wins
per top-level key). That includes coming after values-nodeport.yaml in the NodePort case and
after every other -f in both. --set doesn't follow this rule: Helm always applies --set
after every -f file regardless of command-line position, so a stray --set on
bp-configurator.env here would still win no matter where you put it — step 4 should already
have converted any such override into a values file, not left it inline.
Post-install validation — confirm pods actually come up before declaring success; see
warehouse-<mode>-app/README.md §Post-install validation, but don't run its kubectl get pods -w/port-forward verbatim — those block forever. Use
kubectl wait --for=condition=Ready pod --all -n <namespace> --timeout=5m and a backgrounded
port-forward instead.
Re-run the script whenever NUM_STREAMS or the target GPU changes — the values-override
file isn't tracked automatically; re-generate and re-helm upgrade after a hardware change.
Prerequisites
- Kubernetes cluster reachable via
kubectl, all nodes Ready.
- NVIDIA GPU Operator installed, so nodes report
nvidia.com/gpu as allocatable.
- StorageClass present for VST/Elasticsearch PVCs (
global.storageClass).
- Helm 3.x and kubectl.
- NGC API key for the image pull secret and model/app-data download job.
- Ingress controller installed if using ingress (see the chart README's "No ingress
controller: NodePort" section for the alternative).
- TURN server for WebRTC playback off-cluster (
global.turnServerUrl).
Full detail, values, and exact commands: see
deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app/README.md
§Prerequisites (identical across 2d/3d/mv3dt). This skill only adds the stream-cap step; it
doesn't replace chart setup — the precheck in step 1 is a fast sanity pass, not a substitute for
reading that section on first deploy.
1---2name: vss-deploy-warehouse-helm3description: Use when the user asks to deploy, upgrade, or size the VSS warehouse blueprint (2D / 3D / MV3DT) on Kubernetes via Helm — as opposed to Docker Compose, which is covered by vss-build-vision-ai's warehouse reference. Handles GPU-aware NUM_STREAMS capping so the deployment matches what the perception pipeline can actually sustain.4license: Apache-2.05metadata:6 source: https://skillmd.com/skills/nvidia-ai-blueprints/vss-deploy-warehouse-helm7 version: "3.3.0-rc0"8 author: "NVIDIA Video Search and Summarization team"9 github-url: "https://github.com/NVIDIA-AI-Blueprints/video-search-and-summarization"10 tags: "nvidia blueprint deployment helm kubernetes warehouse"11---1213# VSS Warehouse — Helm Deploy1415## When to Use1617- Deploy, upgrade, or size the VSS warehouse blueprint (2D / 3D / MV3DT) on Kubernetes via Helm18- Compute a GPU-aware `NUM_STREAMS` cap for a warehouse Helm install so it matches what the perception pipeline can sustain1920Do not use this skill for:2122- Docker Compose warehouse deployment — use `vss-build-vision-ai`'s23 [`references/warehouse.md`](../../vss-build-vision-ai/references/profiles/warehouse.md); it owns the24 `HARDWARE_PROFILE` → GPU mapping table and the `blueprint_config.yml` stream-cap semantics25 this skill reuses.26- Non-warehouse Helm profiles (`base`, `search`, `lvs`, `alerts`) — those don't have a27 `bp-configurator` GPU-aware stream cap; deploy them per their own chart READMEs.28- Runtime operations (adding cameras, querying behavior analytics) — use `vss-manage-alerts` /29 `vss-query-analytics` against the running deployment.3031## Why this exists3233Docker Compose's warehouse deploy caps `NUM_STREAMS` per GPU automatically: the configurator reads34`deploy/docker/industry-profiles/warehouse-operations/blueprint-configurator/blueprint_config.yml`'s35`max_streams_supported` table for the detected `HARDWARE_PROFILE` and mode, and clamps36`final_stream_count = min(NUM_STREAMS, max_streams_supported)`.3738The Helm charts (`deploy/helm/industry-profiles/warehouse-operations/warehouse-{2d,3d,mv3dt}-app`)39do **not** do this — their `bp-configurator.env` ships a fixed `NUM_STREAMS` and never sets40`HARDWARE_PROFILE` at all (`ENABLE_PROFILE_CONFIGURATOR=false`). A user who asks for more streams41than the GPU can sustain gets no protection. This skill closes that gap by computing the same cap42Compose would apply and writing it into a Helm values-override file before install.4344## Available Scripts4546| Script | Purpose | Arguments |47|---|---|---|48| [`../../../deploy/helm/industry-profiles/warehouse-operations/scripts/compute_stream_cap.py`](../../../deploy/helm/industry-profiles/warehouse-operations/scripts/compute_stream_cap.py) | Detect GPU (or take an explicit `HARDWARE_PROFILE`), read `max_streams_supported` from `blueprint_config.yml`, cap the requested stream count, and write a `bp-configurator.env`-patched values-override YAML. Pass any values file(s) your install already uses via `-f` so custom `bp-configurator.env` entries in them aren't dropped. | `--mode {2d,3d,mv3dt} --num-streams N [--hardware-profile P] [--gpu-index I] [-f VALUES]... [-o FILE]` |4950This script has no skill/agent dependency — a user who doesn't want to use this skill can run it51directly (`python3 compute_stream_cap.py --mode 2d --num-streams 8`) and pass the generated file to52`helm upgrade/install -f` themselves.5354## Instructions55561. **Precheck the cluster and required inputs** before touching Helm — don't assume a fresh57 cluster already has these. Run each check and report pass/fail back to the user:58 ```bash59 kubectl cluster-info # cluster reachable60 kubectl get nodes # all nodes Ready61 kubectl get storageclass # a StorageClass exists62 kubectl get nodes -o jsonpath='{.items[*].status.allocatable.nvidia\.com/gpu}{"\n"}'63 # non-empty -> GPU Operator has registered GPUs64 helm version --short # Helm 3.x65 ```66 Also ask whether the user already has an NGC API key — that can't be checked from cluster67 state, only asked about.6869 **On any failure, don't just link the user to the README and stop — hand them the actual fix,70 copied from the chart README, and offer to run it for them:**71 - No `StorageClass` → relay the `local-path-provisioner` install + `kubectl patch storageclass`72 snippet from `warehouse-<mode>-app/README.md` §Prerequisites (bare-metal option) — or ask73 what StorageClass they intend to use if they already have one in mind. Multi-node cluster:74 `local-path`'s node affinity can strand `vss-vios-nvstreamer`'s PVCs across different nodes75 (`didn't match PersistentVolume's node affinity`) — relay the same section's76 `nfs-subdir-external-provisioner` snippet instead, and set `vios.vstStorage.vstData`,77 `.vstVideo`, and `.streamerVideos` `.storageClass` to `nfs-client` via three separate `--set`78 flags (or just `global.storageClass`) rather than `local-path`.79 - No `nvidia.com/gpu` allocatable → relay the NVIDIA GPU Operator install steps from80 §Prerequisites (links to the GPU Operator getting-started guide) and the recommended driver81 versions listed there.82 - Cluster unreachable / nodes not `Ready` → this one the user has to fix outside Helm/this83 skill entirely; say so plainly rather than suggesting a chart-level fix.84 - No NGC API key → point at §Required secrets in the chart README for how to create the pull85 secret, don't just say "get an NGC API key."8687 Only proceed to step 2 once cluster/StorageClass/GPU-Operator/Helm all pass and the user has88 confirmed they have an NGC API key — an install started before that will fail partway through89 in a way that's harder to debug than catching it here.902. **Ask ingress vs. NodePort** — this determines both what's installed in this step and which91 install command gets used in step 6, so resolve it before going further, don't default silently92 to one or the other:93 - **Ingress** (needed off-cluster / for a stable hostname) → check whether an ingress94 controller is already installed (`kubectl get ingressclass`). If not, relay the95 `haproxy-ingress` install snippet from `warehouse-<mode>-app/README.md` §"Install the ingress96 controller" and offer to run it. Note this is a one-time, per-cluster step, not per-app.97 - **NodePort** (simplest for a quick local/single-node deploy, no ingress controller needed) →98 tell the user the chart ships `values-nodeport.yaml` for this — the install command in step 699 changes to `-f values-nodeport.yaml` layered under the stream-cap file, and the service URLs100 move to `<NODE_IP>:<port>` instead of `<NODE_IP>/<path>`. See §"No ingress controller:101 NodePort" and §URLs in the chart README for the exact ports.102 If the user hasn't said which they want and there's no clear signal (e.g. "just get it running103 locally" implies NodePort; "expose it for the team" implies Ingress), ask rather than guessing.1043. **Determine mode and whether to enable Alerts:**105 - **Mode.** Use `2d`, `3d`, or `mv3dt` if the request already names one. Otherwise ask —106 don't guess:107 - `2d` — 2D object detection & tracking.108 - `3d` — standalone RTVI-CV-3D / multi-camera 3D tracking on calibrated inputs.109 - `mv3dt` — Multi-View 3D Tracking warehouse profile. Also needs110 `rtvi.vss-rtvi-cv.standaloneWarehouse.mv3dt.fusion.maxExpectedSensors` set to the effective stream111 count in step 6/7 (default `4`) — it's BEV fusion's own camera-count setting, separate from112 `NUM_STREAMS`/`syncFileCount`, and the stream-cap script doesn't touch it.113 - **Alerts.** Not a fourth mode — an optional overlay, off by default, and only available on114 `2d` (`warehouse-2d-app` is the only chart with `vss-alert-bridge`/`agent`/`vss-agent-ui` as115 dependencies; `3d` and `mv3dt` don't have them). If the user is on `3d`/`mv3dt` and asks for116 Alerts, say it's not available there instead of trying to enable it. On `2d`, ask the user117 whether they want it, and explain the tradeoff first rather than enabling or skipping it118 for them: without Alerts they get the raw RT-CV detection/tracking stream; with it, detections119 also pass through a behavior-analytics stage and a VLM verification step (RT-VLM) before120 anything is surfaced as an incident, queryable through the agent/agent UI. That verification121 step is the reason to turn it on — it's what keeps every raw detection from becoming a ticket.122 If they want it, note the four flags have to be set together123 (`vss-alert-bridge.enabled`, `agent.enabled`, `vss-agent-ui.enabled`,124 `rtvi.vss-rtvi-vlm.enabled` — swap the last for an external `vlmBaseUrl` if not using the125 in-cluster VLM) plus Kafka/Elasticsearch/VST endpoint values. Full block:126 `warehouse-2d-app/README.md` §Alerts — layer it in during step 6.127 - **Stream count.** Ask if not given; it sizes the `NUM_STREAMS` cap in step 5.1284. **Ask whether the install customizes `bp-configurator.env`** (extra env vars, different129 defaults) — don't assume none exist just because the user didn't mention one. If they're130 unsure, ask them to check their existing `helm upgrade --install` command for anything touching131 `bp-configurator.env`, file-based or inline. State the outcome back to them either way:132 - **Values file** (`-f my-values.yaml`) → note its path. It gets passed to the script via `-f`133 in the next step *and* to `helm` itself in step 7 — the script's output only carries134 `bp-configurator.env`, so anything else in that file (storage class, ingress, alerts flags)135 still needs `helm` to see the original file directly. See136 [`references/streams.md`](references/streams.md#if-your-install-customizes-bp-configuratorenv).137 - **Inline** (`--set`/`--set-json` on `bp-configurator.env`) → the script only reads YAML files,138 it can't consume a `--set` string. Move it into a values file first — see139 [`references/streams.md`](references/streams.md#if-your-install-customizes-bp-configuratorenv)140 for the `helm get values -a` command (secrets included, handle with care) and why it can't be141 trimmed. Then treat it as the values-file case above.142 - **No customizations** → say so explicitly (e.g. "no custom `bp-configurator.env` overrides,143 so nothing extra is needed here") and proceed without any of the above.1445. **Run the stream-cap script** from the repo root:145 ```bash146 python3 deploy/helm/industry-profiles/warehouse-operations/scripts/compute_stream_cap.py \147 --mode <mode> --num-streams <N> -o values-stream-cap.generated.yaml148 ```149 - If step 4 found a customizing values file, pass it here too via `-f` — otherwise the150 generated file (built from chart defaults, layered last) silently drops those customizations.151 See152 [`references/streams.md`](references/streams.md#if-your-install-customizes-bp-configuratorenv).153 - Without `--hardware-profile`, it runs `nvidia-smi` on GPU index 0 and maps the name to a154 `HARDWARE_PROFILE` using the same table as [`vss-build-vision-ai`'s warehouse155 reference](../../vss-build-vision-ai/references/sizing.md). If detection156 fails or the GPU isn't in that table, pass `--hardware-profile` explicitly. `IGX-THOR`/157 `DGX-SPARK` edge devices aren't supported by this Helm path.158 - No local `nvidia-smi` (running `helm`/`kubectl` from a bastion, laptop, or CI runner rather159 than a GPU node): `kubectl exec` into a GPU Operator daemonset pod (driver or160 device-plugin, e.g. `kubectl get pods --all-namespaces -l app=nvidia-driver-daemonset`) and161 run `nvidia-smi --query-gpu=name --format=csv,noheader` there instead, then map the name and162 pass `--hardware-profile`.163 - It prints the effective (possibly capped) stream count and the `syncFileCount` value to keep164 in step (see [`references/streams.md`](references/streams.md) for why).165 - It never lowers the request silently without saying so — a cap is always logged to stderr.1666. **Prepare the rest of the values** — secrets, storage class, either ingress/`externalHost` or167 the NodePort values file per the choice made in step 2, and — if Alerts was enabled in step 3 —168 the four-flag Alerts values block from `warehouse-2d-app/README.md` §Alerts (Kafka/169 Elasticsearch/VST endpoints included). On `mv3dt`, also add170 `--set rtvi.vss-rtvi-cv.standaloneWarehouse.mv3dt.fusion.maxExpectedSensors=<effective-streams>` (same171 value as `syncFileCount` from step 5). If step 4 found a customizing values file, it goes here172 too (`-f my-values.yaml`) — passing it only to the script in step 5 covers `bp-configurator.env`173 but drops everything else in that file from the install. See174 [`references/streams.md`](references/streams.md) for the full `helm upgrade --install` command175 with the generated file layered in last via `-f`.1767. **Install/upgrade**, chaining the generated file after any other `-f`/`--set` overrides so it177 wins on `bp-configurator.env`. The base command is the same either way; only the178 ingress-vs-NodePort overrides differ:179 ```bash180 helm dependency update deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app181182 # Ingress:183 helm upgrade --install wh deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app \184 -n <namespace> --create-namespace \185 --set global.vssIngress.enabled=true \186 --set global.externalHost=<NODE_IP> \187 --set global.storageClass=<STORAGE_CLASS> \188 --set vios.vss-vios-nvstreamer.syncFileCount=<effective-streams> \189 --set vios.vss-vios-nvstreamer.rtsp.instanceCount=<effective-streams> \190 ... \191 -f values-stream-cap.generated.yaml # last: wins on bp-configurator.env192193 # NodePort:194 helm upgrade --install wh deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app \195 -n <namespace> --create-namespace \196 -f deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app/values-nodeport.yaml \197 --set global.storageClass=<STORAGE_CLASS> \198 --set vios.vss-vios-nvstreamer.syncFileCount=<effective-streams> \199 --set vios.vss-vios-nvstreamer.rtsp.instanceCount=<effective-streams> \200 -f values-stream-cap.generated.yaml # last: wins on bp-configurator.env201 ```202 `...` is the remaining secrets/URL overrides from step 6 — see203 [`references/streams.md`](references/streams.md).204205 `-f values-stream-cap.generated.yaml` has to be the last `-f` in the command — that's what206 makes it win on `bp-configurator.env` (multiple `-f` files merge in order given, later wins207 per top-level key). That includes coming after `values-nodeport.yaml` in the NodePort case and208 after every other `-f` in both. `--set` doesn't follow this rule: Helm always applies `--set`209 after every `-f` file regardless of command-line position, so a stray `--set` on210 `bp-configurator.env` here would still win no matter where you put it — step 4 should already211 have converted any such override into a values file, not left it inline.2128. **Post-install validation** — confirm pods actually come up before declaring success; see213 `warehouse-<mode>-app/README.md` §Post-install validation, but don't run its `kubectl get pods214 -w`/`port-forward` verbatim — those block forever. Use215 `kubectl wait --for=condition=Ready pod --all -n <namespace> --timeout=5m` and a backgrounded216 `port-forward` instead.2179. **Re-run the script whenever `NUM_STREAMS` or the target GPU changes** — the values-override218 file isn't tracked automatically; re-generate and re-`helm upgrade` after a hardware change.219220## Prerequisites221222- **Kubernetes cluster** reachable via `kubectl`, all nodes `Ready`.223- **NVIDIA GPU Operator** installed, so nodes report `nvidia.com/gpu` as allocatable.224- **StorageClass** present for VST/Elasticsearch PVCs (`global.storageClass`).225- **Helm 3.x** and **kubectl**.226- **NGC API key** for the image pull secret and model/app-data download job.227- **Ingress controller** installed if using ingress (see the chart README's "No ingress228 controller: NodePort" section for the alternative).229- **TURN server** for WebRTC playback off-cluster (`global.turnServerUrl`).230231Full detail, values, and exact commands: see232`deploy/helm/industry-profiles/warehouse-operations/warehouse-<mode>-app/README.md`233§Prerequisites (identical across `2d`/`3d`/`mv3dt`). This skill only adds the stream-cap step; it234doesn't replace chart setup — the precheck in step 1 is a fast sanity pass, not a substitute for235reading that section on first deploy.