Skip to content

Commit 65ff6ed

Browse files
akashmjndthaler
akashmjn
and
dthaler
authored
Fix out of date inference deploy configs (#432)
* Update configs to match what is deployed * Increase threshold for Andrews Bay, North SJC, and Port Townsend to 70% * Run all nodes on the f4sv2pool pool * Updated memory config to request 1.6M, with limit 2.5M * Updated images to 01-24-2026.FastAI.R1-12.v0 Signed-off-by: Dave Thaler <dthaler1968@gmail.com> * Update Port Townsend threshold to 70% Per discussion with Akash Signed-off-by: Dave Thaler <dthaler1968@gmail.com> * Copy Strategy Recreate to other configs Since the VMs are memory constrained, stopping the old pod before starting the new one is essential in getting the new one to actually start. Signed-off-by: Dave Thaler <dthaler1968@gmail.com> * Update image on Bush Point and Point Robinson Signed-off-by: Dave Thaler <dthaler1968@gmail.com> * Add comments Signed-off-by: Dave Thaler <dthaler1968@gmail.com> * update for nodes with v1 deployed --------- Signed-off-by: Dave Thaler <dthaler1968@gmail.com> Co-authored-by: Dave Thaler <dthaler1968@gmail.com>
1 parent bd3568d commit 65ff6ed

11 files changed

Lines changed: 119 additions & 31 deletions

‎InferenceSystem/deploy/andrews-bay-configmap.yaml‎

Lines changed: 8 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -5,13 +5,16 @@ metadata:
55
namespace: andrews-bay
66
data:
77
config.yml: |
8-
model_type: "FastAI"
9-
model_local_threshold: 0.5
10-
model_global_threshold: 3
11-
model_path: "./model"
12-
model_name: "model.pkl"
8+
model_id: "orcasound/orcahello-srkw-detector-v1.030926"
9+
model_hf_repo_id: "orcasound/orcahello-srkw-detector-v1"
10+
model_config_path: "./model/config.yaml"
1311
hls_stream_type: "LiveHLS"
1412
hls_polling_interval: 60
1513
hls_hydrophone_id: "rpi_andrews_bay"
1614
upload_to_azure: True
1715
delete_local_wavs: True
16+
model_config_overrides:
17+
inference:
18+
max_batch_size: 2
19+
global_prediction:
20+
pred_global_threshold: 0.6

‎InferenceSystem/deploy/andrews-bay.yaml‎

Lines changed: 14 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,12 @@ metadata:
55
namespace: andrews-bay
66
spec:
77
replicas: 1
8+
# NOTE: This deployment intentionally uses a single replica with a Recreate
9+
# strategy due to memory constraints and so that only one inference pod processes
10+
# hydrophone audio at a time. This avoids concurrent inference instances for
11+
# this site, and the temporary detection gap during upgrades is acceptable.
12+
strategy:
13+
type: Recreate
814
selector:
915
matchLabels:
1016
app: inference-system
@@ -13,13 +19,18 @@ spec:
1319
labels:
1420
app: inference-system
1521
spec:
22+
nodeSelector:
23+
kubernetes.azure.com/agentpool: f4sv2pool
1624
containers:
1725
- name: inference-system
18-
image: orcaconservancycr.azurecr.io/live-inference-system:11-19-2025.FastAI.R1-12.v0
26+
image: orcaconservancycr.azurecr.io/live-inference-system-v1:03-09-2026.v1.0.0
1927
resources:
28+
requests:
29+
cpu: "1"
30+
memory: "1600Mi"
2031
limits:
21-
cpu: 1
22-
memory: 3G
32+
cpu: "1"
33+
memory: "2500Mi"
2334
env:
2435
- name: AZURE_COSMOSDB_PRIMARY_KEY
2536
valueFrom:

‎InferenceSystem/deploy/bush-point.yaml‎

Lines changed: 14 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,12 @@ metadata:
55
namespace: bush-point
66
spec:
77
replicas: 1
8+
# NOTE: This deployment intentionally uses a single replica with a Recreate
9+
# strategy due to memory constraints and so that only one inference pod processes
10+
# hydrophone audio at a time. This avoids concurrent inference instances for
11+
# this site, and the temporary detection gap during upgrades is acceptable.
12+
strategy:
13+
type: Recreate
814
selector:
915
matchLabels:
1016
app: inference-system
@@ -13,13 +19,18 @@ spec:
1319
labels:
1420
app: inference-system
1521
spec:
22+
nodeSelector:
23+
kubernetes.azure.com/agentpool: f4sv2pool
1624
containers:
1725
- name: inference-system
18-
image: orcaconservancycr.azurecr.io/live-inference-system:11-19-2025.FastAI.R1-12.v0
26+
image: orcaconservancycr.azurecr.io/live-inference-system:01-24-2026.FastAI.R1-12.v0
1927
resources:
28+
requests:
29+
cpu: "1"
30+
memory: "1600Mi"
2031
limits:
21-
cpu: 1
22-
memory: 3G
32+
cpu: "1"
33+
memory: "2500Mi"
2334
env:
2435
- name: AZURE_COSMOSDB_PRIMARY_KEY
2536
valueFrom:

‎InferenceSystem/deploy/mast-center.yaml‎

Lines changed: 14 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,12 @@ metadata:
55
namespace: mast-center
66
spec:
77
replicas: 1
8+
# NOTE: This deployment intentionally uses a single replica with a Recreate
9+
# strategy due to memory constraints and so that only one inference pod processes
10+
# hydrophone audio at a time. This avoids concurrent inference instances for
11+
# this site, and the temporary detection gap during upgrades is acceptable.
12+
strategy:
13+
type: Recreate
814
selector:
915
matchLabels:
1016
app: inference-system
@@ -13,13 +19,18 @@ spec:
1319
labels:
1420
app: inference-system
1521
spec:
22+
nodeSelector:
23+
kubernetes.azure.com/agentpool: f4sv2pool
1624
containers:
1725
- name: inference-system
18-
image: orcaconservancycr.azurecr.io/live-inference-system:11-19-2025.FastAI.R1-12.v0
26+
image: orcaconservancycr.azurecr.io/live-inference-system:01-24-2026.FastAI.R1-12.v0
1927
resources:
28+
requests:
29+
cpu: "1"
30+
memory: "1600Mi"
2031
limits:
21-
cpu: 1
22-
memory: 3G
32+
cpu: "1"
33+
memory: "2500Mi"
2334
env:
2435
- name: AZURE_COSMOSDB_PRIMARY_KEY
2536
valueFrom:

‎InferenceSystem/deploy/north-sjc-configmap.yaml‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6,7 +6,7 @@ metadata:
66
data:
77
config.yml: |
88
model_type: "FastAI"
9-
model_local_threshold: 0.5
9+
model_local_threshold: 0.7
1010
model_global_threshold: 3
1111
model_path: "./model"
1212
model_name: "model.pkl"

‎InferenceSystem/deploy/north-sjc.yaml‎

Lines changed: 9 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,12 @@ metadata:
55
namespace: north-sjc
66
spec:
77
replicas: 1
8+
# NOTE: This deployment intentionally uses a single replica with a Recreate
9+
# strategy due to memory constraints and so that only one inference pod processes
10+
# hydrophone audio at a time. This avoids concurrent inference instances for
11+
# this site, and the temporary detection gap during upgrades is acceptable.
12+
strategy:
13+
type: Recreate
814
selector:
915
matchLabels:
1016
app: inference-system
@@ -13,9 +19,11 @@ spec:
1319
labels:
1420
app: inference-system
1521
spec:
22+
nodeSelector:
23+
kubernetes.azure.com/agentpool: f4sv2pool
1624
containers:
1725
- name: inference-system
18-
image: orcaconservancycr.azurecr.io/live-inference-system:11-19-2025.FastAI.R1-12.v0
26+
image: orcaconservancycr.azurecr.io/live-inference-system:01-24-2026.FastAI.R1-12.v0
1927
resources:
2028
limits:
2129
cpu: 1

‎InferenceSystem/deploy/orcasound-lab.yaml‎

Lines changed: 14 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,12 @@ metadata:
55
namespace: orcasound-lab
66
spec:
77
replicas: 1
8+
# NOTE: This deployment intentionally uses a single replica with a Recreate
9+
# strategy due to memory constraints and so that only one inference pod processes
10+
# hydrophone audio at a time. This avoids concurrent inference instances for
11+
# this site, and the temporary detection gap during upgrades is acceptable.
12+
strategy:
13+
type: Recreate
814
selector:
915
matchLabels:
1016
app: inference-system
@@ -13,13 +19,18 @@ spec:
1319
labels:
1420
app: inference-system
1521
spec:
22+
nodeSelector:
23+
kubernetes.azure.com/agentpool: f4sv2pool
1624
containers:
1725
- name: inference-system
18-
image: orcaconservancycr.azurecr.io/live-inference-system:11-19-2025.FastAI.R1-12.v0
26+
image: orcaconservancycr.azurecr.io/live-inference-system:01-24-2026.FastAI.R1-12.v0
1927
resources:
28+
requests:
29+
cpu: "1"
30+
memory: "1600Mi"
2031
limits:
21-
cpu: 1
22-
memory: 3G
32+
cpu: "1"
33+
memory: "2500Mi"
2334
env:
2435
- name: AZURE_COSMOSDB_PRIMARY_KEY
2536
valueFrom:

‎InferenceSystem/deploy/point-robinson.yaml‎

Lines changed: 9 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,12 @@ metadata:
55
namespace: point-robinson
66
spec:
77
replicas: 1
8+
# NOTE: This deployment intentionally uses a single replica with a Recreate
9+
# strategy due to memory constraints and so that only one inference pod processes
10+
# hydrophone audio at a time. This avoids concurrent inference instances for
11+
# this site, and the temporary detection gap during upgrades is acceptable.
12+
strategy:
13+
type: Recreate
814
selector:
915
matchLabels:
1016
app: inference-system
@@ -13,9 +19,11 @@ spec:
1319
labels:
1420
app: inference-system
1521
spec:
22+
nodeSelector:
23+
kubernetes.azure.com/agentpool: f4sv2pool
1624
containers:
1725
- name: inference-system
18-
image: orcaconservancycr.azurecr.io/live-inference-system:11-19-2025.FastAI.R1-12.v0
26+
image: orcaconservancycr.azurecr.io/live-inference-system:01-24-2026.FastAI.R1-12.v0
1927
resources:
2028
limits:
2129
cpu: 1

‎InferenceSystem/deploy/port-townsend-configmap.yaml‎

Lines changed: 8 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -5,13 +5,16 @@ metadata:
55
namespace: port-townsend
66
data:
77
config.yml: |
8-
model_type: "FastAI"
9-
model_local_threshold: 0.5
10-
model_global_threshold: 3
11-
model_path: "./model"
12-
model_name: "model.pkl"
8+
model_id: "orcasound/orcahello-srkw-detector-v1.030926"
9+
model_hf_repo_id: "orcasound/orcahello-srkw-detector-v1"
10+
model_config_path: "./model/config.yaml"
1311
hls_stream_type: "LiveHLS"
1412
hls_polling_interval: 60
1513
hls_hydrophone_id: "rpi_port_townsend"
1614
upload_to_azure: True
1715
delete_local_wavs: True
16+
model_config_overrides:
17+
inference:
18+
max_batch_size: 2
19+
global_prediction:
20+
pred_global_threshold: 0.6

‎InferenceSystem/deploy/port-townsend.yaml‎

Lines changed: 14 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,12 @@ metadata:
55
namespace: port-townsend
66
spec:
77
replicas: 1
8+
# NOTE: This deployment intentionally uses a single replica with a Recreate
9+
# strategy due to memory constraints and so that only one inference pod processes
10+
# hydrophone audio at a time. This avoids concurrent inference instances for
11+
# this site, and the temporary detection gap during upgrades is acceptable.
12+
strategy:
13+
type: Recreate
814
selector:
915
matchLabels:
1016
app: inference-system
@@ -13,13 +19,18 @@ spec:
1319
labels:
1420
app: inference-system
1521
spec:
22+
nodeSelector:
23+
kubernetes.azure.com/agentpool: f4sv2pool
1624
containers:
1725
- name: inference-system
18-
image: orcaconservancycr.azurecr.io/live-inference-system:11-19-2025.FastAI.R1-12.v0
26+
image: orcaconservancycr.azurecr.io/live-inference-system-v1:03-09-2026.v1.0.0
1927
resources:
28+
requests:
29+
cpu: "1"
30+
memory: "1600Mi"
2031
limits:
21-
cpu: 1
22-
memory: 3G
32+
cpu: "1"
33+
memory: "2500Mi"
2334
env:
2435
- name: AZURE_COSMOSDB_PRIMARY_KEY
2536
valueFrom:

0 commit comments

Comments
 (0)