OVDC: Helm Values#
values.yaml has two halves, and the distinction matters:
settings.*mirrors the application’s own config schema (rust/ovdc/src/config.tomlin the OVDC repository) and is rendered key-by-key into the ConfigMap’sconfig.toml. Every key here must be one the binary deserializes; a missing required key panics the process at startup.Everything else is Kubernetes-level chart configuration.
The container re-reads config.toml every 5 seconds and applies
changes without a restart, but only for settings that are read per-use:
logFormat / logLevel / logFilter, all otel* /
telemetry sections (including otelTraces.level / filter and
otelLogs.level / filter), garbageCollection, and
store.blobCutoff / blobChunkSize. store.maxSize,
store.mode, the whole engine section, grpc, grpcTls, and
grpcJwt are read once at startup, so changing them requires a pod
restart. A reloaded file that fails validation panics the pod, so a bad
helm upgrade surfaces as CrashLoopBackOff.
Top-Level Field Reference#
image.repository— string, defaultnvidia/omniverse/ovderivedcache. Image repository underimage.registry.image.pullSecrets— list of{name}, default[{name: regcred}]. KubernetesimagePullSecretsentries.nameOverride,fullnameOverride— string, default"". Override generated resource-name components. See the deployment guide for the naming pattern.replicas— integer, default1. Number of OVDC pods. Cached entries are valid for the count used to populate them. Increasing or decreasing the count invalidates entries for the previous count while a different count is active; returning to that count makes its retained entries valid again. Plan for warm-up of data not already cached for the active count; see Provision and Scale.podSecurityContext,securityContext— Kubernetes securityContext object, default non-root uid/gid 1000 with all capabilities dropped. Pod- and container-level security settings.resources— Kubernetes resources object, defaultrequests.memory: 56G. Passed through verbatim to the container.storage.volume.enabled— boolean, defaulttrue. Whenfalse, the DB is ephemeral (container filesystem)—dev/test only. Not hot-reloaded (immutable).storage.volume.size,storageClassName— quantity string / string, default300Gi/managed-csi. Per-pod PersistentVolumeClaim sizing and class. Not hot-reloaded (volumeClaimTemplatesare immutable).presets.cpu,presets.ram— integer enum (1-8) / string enum, default~(unset). Optional sizing presets that override a fixed table ofsettings.enginevalues at render time; see Configure for the tables. Unset leavessettings.engineas written. CPU presets above4also setsettings.engine.db.max_subcompactionsto4. Not hot-reloaded (pod restart, same assettings.engine).settings.store.*— object, defaultmaxSize: AUTO,blobCutoff: 2500,blobChunkSize: 4000000,mode: "log". RocksDB storage sizing and mode; see Configure.maxSizeandmodeare not hot-reloaded;blobCutoffandblobChunkSizeare.settings.engine.*— object, defaultcacheSize: "32G",blockCacheSize: "8G", plus RocksDBdb,cf, andblockoption maps. RocksDB engine tuning:cacheSize,blockCacheSize, and thedb,cf,block,blobsCf,blobsBlockoption maps.cacheSize,blockCacheSize, andblobsCf.max_write_buffer_numberare overridden bypresets.ramwhen it is set;increaseParallelism,threadsHigh, andthreadsLoware overridden bypresets.cpu. Not hot-reloaded (pod restart).settings.garbageCollection.*— object, defaultminFreeCapacity: 40,deleteKeyspaceQuantile: 60,cooldownSecs: 10. Garbage collection thresholds and the pause between passes; the capacity check itself runs on a fixed 5-second interval. Hot-reloaded.settings.telemetry.*— object, defaultprometheusMetricsExposition: true,prometheusMetricsPort: 3051. Prometheus metrics exposition. Hot-reloaded.settings.logFilter— string, default"". Optional per-module level overrides for the stdout sink, layered on top oflogLevel(format:module[*]=level[,module2[*]=level2]..., for example"ovdc::grpc=debug"). Independent ofotelTraces.level/filterandotelLogs.level/filter. Hot-reloaded.settings.otel,otelTraces,otelLogs,otelMetrics— object, defaultexporter: "otlp",endpoint: "http://localhost:4317". OpenTelemetry export configuration, one section per signal.otelTraces.level/otelLogs.levelset the base level for that sink (enum, same values aslogLevel);otelTraces.filter/otelLogs.filterare per-module overrides only (same format aslogFilter), independent oflogLeveland of each other. Hot-reloaded.settings.grpc.*— object, defaultport: 3010plus per-field defaults documented inline below. gRPC service tuning: timeouts, keepalives, message size limits. Not hot-reloaded (pod restart).settings.grpc.tlsHandshakeTimeout— positive duration, default"10s". Deadline for completing a TLS handshake; expiry closes the socket and incrementsovdc_api_tls_handshake_timeout_total. Read at startup; changes require a pod restart.settings.grpcTls.*— object, defaultenabled: false. TLS for the gRPC listener; see TLS Configuration. Not hot-reloaded (pod restart).settings.grpcJwt.*— object, defaultenabled: false. JWT verification for the gRPC listener. Not hot-reloaded (pod restart).service.*— object, defaultloadBalancer: false. Service ports are not configurable here—they derive fromsettings.grpc.portandsettings.telemetry.prometheusMetricsPort.prometheusAnnotation— boolean, defaulttrue. Addsprometheus.io/*scrape annotations to the metrics Service.debug— boolean, defaultfalse. Whentrue, replaces the container command withsleep—development only, never enable in production.
The full field-level YAML, including every RocksDB engine.db /
cf / block option and every grpc / grpcTls / grpcJwt
/ otel* key with inline descriptions, follows below.
# Labels added to the StatefulSet, its volume claims and the pod template.
global:
additionalLabels:
image:
# Registry and repository the release image is pulled from.
registry: nvcr.io
repository: nvidia/omniverse/ovderivedcache
pullPolicy: IfNotPresent
# Rendered as the pod's imagePullSecrets, so each entry is a `name:` mapping
# referencing a kubernetes.io/dockerconfigjson Secret in this namespace.
pullSecrets:
- name: regcred
# Image tag, used only when overrideTag is true. Otherwise the chart's
# appVersion is used.
tag: "latest"
overrideTag: false
# Override the chart name used to build resource names. The release name still
# prefixes them: release `foo` with nameOverride `bar` yields `foo-bar`.
nameOverride: ""
# Replace the generated resource name outright, ignoring the release name.
fullnameOverride: ""
# The number of pods to deploy.
replicas: 1
# Node label constraints for the pods, e.g. `agentpool: ovdcpool`.
nodeSelector: {}
# Taints the pods tolerate, in the standard Kubernetes tolerations format.
tolerations: []
# When true, the scheduler prefers to place ovdc pods on nodes that do not
# already run one. Merged into `affinity` below; values set there win on
# conflict. Set false to drop the chart's rule entirely.
selfAntiAffinity: true
# Additional affinity rules for the pod template.
affinity: {}
# Annotations added to the pod template. Changing these rolls the StatefulSet.
podAnnotations: {}
# Pod-level securityContext, applied to every container in the pod. fsGroup
# makes the non-root default work by chowning the mounted volume at /env.
podSecurityContext:
runAsNonRoot: true
runAsUser: 1000
runAsGroup: 1000
fsGroup: 1000
seccompProfile:
type: RuntimeDefault
# Passed to the container as RUST_LOG and RUST_BACKTRACE. RUST_LOG is only a
# fallback for the stdout log level: settings.logLevel wins whenever it parses.
rustLog: "info"
rustBacktrace: "full"
# Container-level securityContext for the ovdc container.
securityContext:
runAsNonRoot: true
runAsUser: 1000
runAsGroup: 1000
allowPrivilegeEscalation: false
capabilities:
drop:
- ALL
# Compute resources for the ovdc container, passed through verbatim. Should be
# at least engine.cacheSize + engine.blockCacheSize + ~2-6GB network buffers.
# Write buffers share the block cache when useWriteBufferManager is true;
# otherwise add both column families' write buffers separately.
resources:
requests:
memory: 56G
#limits:
# memory: 248G
storage:
volume:
# When enabled the DB environment resides on a k8s mounted volume. When
# false the DB environment is ephemeral (container filesystem)—dev/test only.
enabled: true
# Size of the requested PersistentVolumeClaim. Drives settings.store.maxSize
# when that value is "AUTO". volumeClaimTemplates are immutable—changing
# this on an existing release is rejected on upgrade.
size: 300Gi
# Azure-specific default; change to match a class available in your cluster.
# Must support ReadWriteOnce; each replica gets its own volume.
storageClassName: managed-csi
# Convenience sizing presets. When set, each one overrides a fixed set of
# settings.engine values below at render time (see Configure's Sizing Presets
# section for the tables). Leave unset (null) to use settings.engine as
# written instead. Any other value is rejected by the schema before the
# chart renders.
presets:
# One of 1, 2, 3, 4, 5, 6, 7, 8. Maps to settings.engine.increaseParallelism
# / threadsHigh / threadsLow. CPU presets above 4 also set
# settings.engine.db.max_subcompactions to 4.
cpu: ~
# One of "4GB", "6GB", "7GB", "8GB", "12GB", "16GB", "24GB", "32GB", "48GB",
# "64GB". Maps to settings.engine.cacheSize / blockCacheSize and
# settings.engine.blobsCf.max_write_buffer_number.
ram: ~
settings:
# Allowed values: json, human, human_no_color
logFormat: "json"
# Log level for stdout. Allowed values: error, warn, info, debug, trace, off.
logLevel: "info"
# Optional per-module level overrides for the stdout sink, layered on top of
# logLevel. Independent of otelTraces.level/filter and otelLogs.level/filter.
# Format: module[*]=level[,module2[*]=level2]... e.g. "ovdc::grpc=debug"
logFilter: ""
store:
# Target max size for the application to use ("AUTO" derives it from
# storage.volume.size: the whole volume less min(20%, 100GB) headroom).
# Requires storage.volume.enabled: true, otherwise set an explicit size
# such as "275G".
maxSize: AUTO
# Values larger than this are written as a chunked blob.
blobCutoff: 2500
# Size by which blob values above blobCutoff are split.
blobChunkSize: 4000000
# Storage engine mode. Allowed values: legacy, legacy_and_discrete, log.
# Changing this on a populated database changes how values are read back.
mode: "log"
# RocksDB engine configuration (rust/ovdc_map/src/environment/rocks_adapter/config.rs).
# Read once at startup; changes require a pod restart.
engine:
# When true, memtable memory is charged against the same LRU cache as
# blockCacheSize.
useWriteBufferManager: true
# Size of the in-process row cache. This is OVDC's own cache, not RocksDB's. 0 disables it.
cacheSize: "32G"
# Size of the RocksDB LRU block cache.
blockCacheSize: "8G"
# RocksDB increase_parallelism: total background threads for flushes/compactions.
increaseParallelism: 8
# Env high/low priority background thread counts.
threadsHigh: 4
threadsLow: 4
# Passed to RocksDB verbatim, merged over the binary's built-in defaults.
# db: DBOptions (database-wide). cf/block: CFOptions/BlockBasedTableOptions
# for the 'default' column family.
db:
allow_concurrent_memtable_write: true
max_open_files: 128
max_background_jobs: 6
table_cache_numshardbits: 8
fail_if_options_file_error: true
keep_log_file_num: 8
avoid_unnecessary_blocking_io: true
max_log_file_size: "64M"
stats_dump_period_sec: 0
create_if_missing: true
max_subcompactions: 2
cf:
max_write_buffer_number: 4
min_write_buffer_number_to_merge: 1
write_buffer_size: "64M"
level0_stop_writes_trigger: 24
blob_garbage_collection_force_threshold: 0.25
min_blob_size: 8096
enable_blob_files: true
enable_blob_garbage_collection: true
memtable_whole_key_filtering: true
periodic_compaction_seconds: 28800
memtable_prefix_bloom_size_ratio: 0.020000
block:
cache_index_and_filter_blocks: true
block_size: 16384
index_type: "kTwoLevelIndexSearch"
filter_policy: "bloomfilter:10"
partition_filters: true
optimize_filters_for_memory: true
data_block_index_type: "kDataBlockBinaryAndHash"
pin_l0_filter_and_index_blocks_in_cache: true
prepopulate_block_cache: "kFlushOnly"
# Options for the 'blobs' column family (exists when store.mode is "log" or
# "legacy_and_discrete"). Tuned as RocksDB's integrated BlobDB storage with
# aggressive garbage collection, plus write-buffer/compaction settings
# cloned from 'default' above so it isn't left on RocksDB's much smaller
# stock defaults.
blobsCf:
max_write_buffer_number: 24
min_write_buffer_number_to_merge: 1
write_buffer_size: "64M"
level0_stop_writes_trigger: 24
blob_garbage_collection_force_threshold: 0.25
enable_blob_files: true
min_blob_size: 0
blob_file_size: "64M"
enable_blob_garbage_collection: true
blob_garbage_collection_age_cutoff: 0.8
blob_compaction_readahead_size: "4M"
memtable_whole_key_filtering: true
periodic_compaction_seconds: 28800
memtable_prefix_bloom_size_ratio: 0.020000
blobsBlock:
cache_index_and_filter_blocks: true
block_size: 16384
index_type: "kTwoLevelIndexSearch"
filter_policy: "bloomfilter:10"
partition_filters: true
optimize_filters_for_memory: true
data_block_index_type: "kDataBlockBinaryAndHash"
pin_l0_filter_and_index_blocks_in_cache: true
prepopulate_block_cache: "kFlushOnly"
# Garbage collection configuration. A background task compares DB usage
# against settings.store.maxSize every 5 seconds.
garbageCollection:
# If free map space falls below this percentage, GC begins.
minFreeCapacity: 40
# Approximate percentage of live keys GC attempts to remove once it begins.
deleteKeyspaceQuantile: 60
# How long to wait after a GC pass finishes before checking free capacity again.
cooldownSecs: 10
# HTTP metrics service serving Prometheus plain text metrics.
telemetry:
prometheusMetricsExposition: true
prometheusMetricsPort: 3051
# OpenTelemetry integration. otel.enabled/useEnvVars are required by the
# schema but not independently actionable—disable export per-signal below.
otel:
enabled: true
useEnvVars: true
otelTraces:
exporter: "otlp"
otlp:
protocol: "grpc"
endpoint: "http://localhost:4317"
timeout: "10s"
processor:
scheduleDelay: 5000
maxQueueSize: 2048
maxBatchSize: 512
attributeValueLengthLimit: "4K"
attributeCountLimit: 128
eventCountLimit: 128
linkCountLimit: 128
eventAttributeCountLimit: 128
linkAttributeCountLimit: 128
samplerArg: 0.25
sampler: "always_on"
# Base level for events forwarded to the span exporter. Independent of
# logLevel and otelLogs.level. Allowed values: error, warn, info, debug, trace, off.
level: "info"
# Per-module level overrides layered on top of `level`.
# Format: module[*]=level[,module2[*]=level2]...
filter: ""
otelLogs:
exporter: "otlp"
otlp:
protocol: "grpc"
endpoint: "http://localhost:4317"
timeout: "10s"
processor:
scheduleDelay: 5000
maxQueueSize: 2048
maxBatchSize: 512
attributeValueLengthLimit: "4K"
attributeCountLimit: 128
# Base level for events forwarded to the log exporter. Independent of
# logLevel and otelTraces.level. Allowed values: error, warn, info, debug, trace, off.
level: "info"
# Per-module level overrides layered on top of `level`.
# Format: module[*]=level[,module2[*]=level2]...
filter: ""
otelMetrics:
exporter: "otlp"
otlp:
protocol: "grpc"
endpoint: "http://localhost:4317"
timeout: "15s"
processor:
interval: "30s"
filter: "info"
# gRPC HTTP service. Read once at startup; changes require a pod restart.
grpc:
# Port for the gRPC service. Also used for the container port, both
# Services' targetPort, and the readiness/liveness probes.
port: 3010
initialStreamWindowSize: "512K"
initialConnectionWindowSize: "32M"
connectionConcurrencyLimit: 32
maxConcurrentStreams: 0
# Positive TLS handshake deadline. Read at startup; changes require a pod restart.
tlsHandshakeTimeout: "10s"
maxFrameSize: 0
timeout: "30s"
writeTimeout: "15s"
tcpKeepaliveAfterIdleTime: "3s"
tcpKeepaliveInterval: "3s"
tcpKeepaliveRetries: 5
http2KeepaliveInterval: "3s"
http2KeepaliveTimeout: "20s"
# If true a tenant ID can be set through a string in the metadata (unauthenticated).
tenantFromMetadata: true
maxDecodingMessageSize: "5M"
# Max time a single PUT stream can run for. Capped at 4h.
putTimeout: "1h"
# TLS for gRPC. Certificate/key paths are fixed by the chart at /etc/ovdc/tls
# and are not configurable—only secretName and includeCaRoot are.
grpcTls:
enabled: false
secretName: ""
includeCaRoot: false
# JWT verification for gRPC.
grpcJwt:
enabled: false
require: false
jwkPublicKeysetUrl: []
jwkUpdateIntervalSecs: 60
cacheSizeMb: 16
verifyAud: false
audClaim: ""
verifyExp: false
# Service ports are not configurable here—they are derived from
# settings.grpc.port and settings.telemetry.prometheusMetricsPort.
service:
annotations: {}
# One LoadBalancer Service per pod (named <fullname>-<ordinal>), allowing an
# external IP per replica. The headless Service is always created.
loadBalancer: false
loadBalancerSourceRanges: []
loadBalancerIP: ""
# Adds prometheus.io scrape annotations to the metrics service.
prometheusAnnotation: true
# Values for the prometheus.io scrape annotations above.
monitoring:
path: /metrics
scheme: http
# When true the ovdc container's command is replaced with `sleep` (service does
# not run) and an extra <fullname>-debug pod is created with config.toml mounted
# at /config for inspection. Development only.
debug: false