Kubernetes Dependency Reduction Assessment v3 (Fresh Analysis)

Date: December 22, 2025 Codebase: k8s.io/kubernetes (master branch) Commit: 9720186a466cc627b0417433fb1f66cd1dd96f94 Vendor Size: 68 MB (4,349 Go files across 1,215 directories)

Executive Summary

This is a completely fresh, blind analysis - discovered independently without referencing prior assessments. The analysis identifies:

Zero-usage packages - vendored but never imported by k8s code
Single-use packages - imported in only 1 file
Transitive-only chains - packages only used by other vendor code
Duplicate libraries - multiple packages serving the same purpose
Inlinable micro-packages - tiny utilities that could be copied

Key Fresh Findings

Category	Discovery	Estimated Savings
etcd server transitive chain	6+ packages only used by etcd/server	~600 KB
Semver duplication	3 libraries, only 1 actually used	2 redundant packages
Antlr4 (CEL parser)	560 KB, vendor-only usage	Unavoidable (CEL dep)
Smallest inlinable packages	5 packages under 200 LOC	5 deps removable
Protobuf triplication	3 implementations required	Unavoidable

Part 1: Packages with Zero Non-Vendor Usage

These packages exist in vendor but are never imported by any Kubernetes-owned code:

etcd Server Transitive Chain

These packages are ONLY imported by go.etcd.io/etcd/server/v3/*:

Package	Size	Imported By
`github.com/tmc/grpc-websocket-proxy`	20 KB	etcd/server/embed
`github.com/xiang90/probing`	24 KB	etcd/server/rafthttp
`github.com/soheilhy/cmux`	56 KB	etcd/server/embed
`github.com/dustin/go-humanize`	64 KB	etcd/server (10 files)
`github.com/jonboulle/clockwork`	60 KB	etcd/server (8 files)
`github.com/golang-jwt/jwt`	160 KB	etcd/server/auth

Total: ~384 KB of packages only used by etcd server internals

Other Vendor-Only Packages

Package	Size	Imported By
`github.com/cenkalti/backoff`	44 KB	OTEL retry only
`github.com/NYTimes/gziphandler`	56 KB	kube-openapi handler only
`github.com/antlr4-go/antlr`	560 KB	cel-go parser only
`github.com/kylelemons/godebug`	?	prometheus testutil only
`github.com/coreos/go-semver`	28 KB	etcd only
`github.com/Masterminds/semver`	36 KB	ginkgo internal only

Part 2: Packages with Single Non-Vendor Import

Package	LOC	Single Use Location
`github.com/armon/go-socks5`	~200	`apimachinery/pkg/util/httpstream/spdy/roundtripper_test.go`
`github.com/fatih/camelcase`	90	`kubectl/pkg/describe/describe.go`
`github.com/beorn7/perks/quantile`	?	prometheus summary (vendor)
`github.com/felixge/httpsnoop`	100	vendor only
`github.com/JeffAshton/win_pdh`	?	`kubelet/winstats/perfcounters.go` (Windows only)

Part 3: Smallest Packages (Inlining Candidates)

Discovered by scanning all vendor packages for LOC:

Package	Total LOC	Files	Complexity
`github.com/josharian/intern`	44	1	Trivial string interning
`github.com/lithammer/dedent`	49	1	Regex whitespace removal
`github.com/inconshreveable/mousetrap`	58	2	Windows trap detection
`github.com/mitchellh/go-wordwrap`	83	1	Word-boundary line wrapping
`github.com/fatih/camelcase`	90	1	CamelCase splitting
`github.com/MakeNowJust/heredoc`	105	1	Here-doc processing
`github.com/mohae/deepcopy`	125	1	Deep copy via reflection
`github.com/munnerz/goautoneg`	189	1	HTTP content negotiation
`github.com/stoewer/go-strcase`	191	5	Case conversion

Usage Analysis for Smallest Packages

Package	Non-Vendor Imports	Recommendation
`josharian/intern`	0 (easyjson only)	Keep (transitive)
`lithammer/dedent`	44 files	INLINE - heavily used
`inconshreveable/mousetrap`	0 (cobra only)	Keep (transitive)
`mitchellh/go-wordwrap`	2 files	INLINE - kubectl only
`fatih/camelcase`	1 file	INLINE - single use
`MakeNowJust/heredoc`	1 file	INLINE - single use
`mohae/deepcopy`	0 (vendor only)	Keep (transitive)
`munnerz/goautoneg`	3 files	Consider inlining
`stoewer/go-strcase`	0 (cel-go only)	Keep (transitive)

Part 4: Duplicate/Overlapping Libraries

Semver Libraries (3 implementations!)

Library	Size	Non-Vendor Usage
`github.com/blang/semver/v4`	56 KB	28 files - actual usage
`github.com/coreos/go-semver`	28 KB	0 files - etcd transitive
`github.com/Masterminds/semver/v3`	36 KB	0 files - ginkgo transitive

Finding: Only blang/semver is used by Kubernetes. The other two are transitive dependencies.

Protobuf Libraries (3 implementations)

Library	Size	Purpose
`google.golang.org/protobuf`	1.4 MB	Modern official
`github.com/golang/protobuf`	152 KB	Legacy shim
`github.com/gogo/protobuf`	1.3 MB	Performance (etcd, prometheus, containerd)

Finding: All three are required. gogo/protobuf is needed by etcd, prometheus, and containerd APIs.

Logging Libraries

Library	Size	Non-Vendor Usage
`github.com/sirupsen/logrus`	160 KB	25 files
`go.uber.org/zap`	600 KB	~12 files (via zapr)
`github.com/go-logr/logr`	216 KB	61 files

Finding: Multiple logging abstractions coexist. klog/logr is primary, zap is for JSON output.

Part 5: Largest Transitive Dependency Chains

Chain 1: etcd Server (1.9 MB) - TEST ONLY

go.etcd.io/etcd/server/v3 (1.9 MB, 231 files)
  ├── Brings in: bbolt (388 KB)
  ├── Brings in: raft (472 KB)
  ├── Brings in: grpc-websocket-proxy (20 KB)
  ├── Brings in: probing (24 KB)
  ├── Brings in: cmux (56 KB)
  ├── Brings in: go-humanize (64 KB)
  ├── Brings in: clockwork (60 KB)
  ├── Brings in: golang-jwt (160 KB)
  └── Brings in: coreos/go-semver (28 KB)

Critical Finding: etcd/server is only imported by 2 test files:

staging/src/k8s.io/apiserver/pkg/storage/etcd3/testserver/test_server.go
staging/src/k8s.io/apiserver/pkg/storage/etcd3/store_test.go

Total transitive bloat from etcd/server: ~3.2 MB

Chain 2: CEL/Antlr (560 KB)

github.com/google/cel-go (2.0 MB)
  └── Brings in: antlr4-go/antlr (560 KB)

Finding: antlr4 is required for CEL expression parsing. No alternative.

Chain 3: OTEL Tracing

go.opentelemetry.io/otel (3.2 MB)
  └── Brings in: cenkalti/backoff (44 KB) - retry logic

Part 6: Top Import Counts (Most-Used Packages)

From blind analysis of all non-vendor imports:

Package	Import Count	Notes
`github.com/google/go-cmp/cmp`	509	Test comparisons
`github.com/onsi/ginkgo/v2`	455	BDD testing
`github.com/stretchr/testify/assert`	384	Assertions
`github.com/onsi/gomega`	330	Matchers
`github.com/stretchr/testify/require`	270	Required assertions
`github.com/spf13/cobra`	239	CLI framework
`github.com/spf13/pflag`	216	Flags
`github.com/google/cel-go/common/types/ref`	125	CEL types
`github.com/google/cadvisor/info/v1`	102	Container metrics
`github.com/google/cel-go/cel`	81	CEL API
`github.com/prometheus/client_golang/prometheus`	69	Metrics

Part 7: Packages Used in Only 2-3 Files

Package	File Count	Files
`github.com/mitchellh/go-wordwrap`	2	kubectl help_flags_printer, term_writer
`github.com/peterbourgon/diskv`	2	client-go disk cache
`github.com/coredns/corefile-migration`	2	kubeadm dns migration
`github.com/jonboulle/clockwork`	2	kubectl diff, patcher (non-etcd)
`github.com/ishidawataru/sctp`	3	agnhost test images only
`github.com/exponent-io/jsonpath`	2	kubectl schema, patch_test
`github.com/robfig/cron`	3	cronjob controller

Part 8: Size Analysis

Top 15 Largest github.com Directories

Directory	Size
`github.com/google/`	4.8 MB
`github.com/onsi/`	1.5 MB
`github.com/prometheus/`	1.4 MB
`github.com/gogo/protobuf/`	1.3 MB
`github.com/vishvananda/`	1.0 MB
`github.com/containerd/`	720 KB
`github.com/grpc-ecosystem/`	692 KB
`github.com/opencontainers/`	680 KB
`github.com/spf13/`	632 KB
`github.com/stretchr/`	568 KB
`github.com/antlr4-go/`	560 KB
`github.com/Microsoft/`	520 KB
`github.com/moby/`	356 KB
`github.com/fxamacker/`	352 KB
`github.com/json-iterator/`	344 KB

Top-Level Vendor Breakdown

Directory	Size
`golang.org/x/`	23 MB
`github.com/`	21 MB
`sigs.k8s.io/`	7.0 MB
`go.etcd.io/`	4.7 MB
`google.golang.org/`	4.0 MB
`go.opentelemetry.io/`	3.7 MB
`k8s.io/`	2.2 MB

Part 9: Optimization Recommendations

Tier 1: Quick Wins (Inline Small Packages)

Action	Savings	Effort
Inline `lithammer/dedent` (49 LOC)	1 dep	1 hour
Inline `fatih/camelcase` (90 LOC)	1 dep	1 hour
Inline `mitchellh/go-wordwrap` (83 LOC)	1 dep	1 hour
Inline `MakeNowJust/heredoc` (105 LOC)	1 dep	1 hour

Total: 4 dependencies, ~4 hours work

Tier 2: Investigate etcd Server Isolation

The entire etcd/server module (1.9 MB + ~1.3 MB transitive) is only used for test infrastructure.

Options:

Move test server to separate test-only module
Use build tags to exclude from production
Accept as test infrastructure cost

Potential savings: 3.2 MB vendor, significant binary reduction

Tier 3: Accept as Necessary

Package	Reason to Keep
`antlr4-go/antlr`	Required by CEL parser
`gogo/protobuf`	Required by etcd, prometheus, containerd
`coreos/go-semver`	etcd transitive (unavoidable with etcd)
`Masterminds/semver`	ginkgo transitive (test framework)

Part 10: Packages by Primary Consumer

Only Used by Kubernetes Tests

Package	Primary Consumer
`github.com/onsi/ginkgo/v2`	e2e tests
`github.com/onsi/gomega`	e2e tests
`github.com/stretchr/testify`	unit tests
`github.com/armon/go-socks5`	roundtripper_test.go

Only Used by Specific Components

Package	Component
`github.com/coredns/`	kubeadm only
`github.com/chai2010/gettext-go`	kubectl i18n only
`github.com/JeffAshton/win_pdh`	kubelet Windows only
`github.com/ishidawataru/sctp`	agnhost test images only
`github.com/robfig/cron`	cronjob controller only
`github.com/russross/blackfriday`	kubectl markdown help

Conclusion

Fresh Findings Summary

etcd/server is the biggest opportunity - 3.2 MB of vendor brought in for 2 test files
4 micro-packages can be trivially inlined - dedent, camelcase, wordwrap, heredoc (total ~330 LOC)
3 semver libraries exist, only 1 is used - blang/semver is the real one; coreos and Masterminds are transitive
3 protobuf implementations are required - ecosystem dependencies prevent consolidation
antlr4 (560 KB) is unavoidable - required for CEL expression parsing
~384 KB of packages are only used by etcd/server internals - would be removed if etcd/server were isolated

Actionable Items

Priority	Action	Impact
1	Inline 4 micro-packages	-4 deps, trivial
2	Investigate etcd server test isolation	-3.2 MB potential
3	Document why 3 protobuf impls needed	Prevents future confusion
4	Accept CEL/antlr cost	Unavoidable for validation

Freshly discovered through blind analysis of k8s.io/kubernetes vendor directory Analysis performed: December 22, 2025

Appendix: Prompt for Regenerating This Analysis

Use the following prompt with Claude Code to regenerate this analysis on a newer commit:

Perform a fresh, blind analysis of the Kubernetes vendor directory to identify dependency reduction opportunities. Do NOT reference any prior analysis - discover everything independently.

Analyze and report on:

1. **Vendor overview**: Total size, file counts, directory structure breakdown

2. **Zero-usage packages**: Find vendored packages that are NEVER imported by any k8s.io/* code (only by other vendor code). Use grep to verify import counts.

3. **Single-use packages**: Packages imported by only 1-2 non-vendor files

4. **Transitive dependency chains**: Map which large packages (etcd/server, CEL, OTEL) bring in which smaller packages that wouldn't otherwise be needed

5. **Smallest packages by LOC**: Find packages under 200 lines that could potentially be inlined

6. **Duplicate libraries**: Identify multiple libraries serving the same purpose (semver, logging, protobuf, etc.)

7. **Top imported packages**: Count imports to find most/least used dependencies

For each finding, provide:
- Package path and size (use `du -sh`)
- Actual import count (use grep on non-vendor .go files)
- What imports it (vendor-only vs k8s code)
- Actionable recommendation

Output as a markdown report with tables, organized by category. Include the git commit hash for reproducibility.

Be thorough - scan the actual vendor directory, count real imports, verify claims with grep. This should take 15-20 minutes of analysis.

dims/kubernetes-dependency-analysis.md

Select an option

No results found