Nit: chaining two (*logrus.Entry).WithField calls is surprisingly inefficient compared to .WithFields as it is implemented in terms of the latter and deep-copies the Entry on each call.

goos: darwin
goarch: amd64
pkg: github.com/corhere/logbench
cpu: Intel(R) Core(TM) i9-8950HK CPU @ 2.90GHz
BenchmarkWithField-12     	   10000	    344272 ns/op	  338640 B/op	      15 allocs/op
BenchmarkWithFields-12    	 1000000	      2372 ns/op	     566 B/op	       3 allocs/op
PASS

package logbench

import (
	"io"
	"strconv"
	"testing"

	"github.com/sirupsen/logrus"
)

func init() {
	logrus.SetOutput(io.Discard)
}

func BenchmarkWithField(b *testing.B) {
	entry := logrus.NewEntry(logrus.StandardLogger())
	b.ResetTimer()
	b.ReportAllocs()
	for i := 0; i < b.N; i++ {
		entry = entry.WithField(strconv.Itoa(i), i)
	}
	entry.Info("hi")
}

func BenchmarkWithFields(b *testing.B) {
	entry := logrus.NewEntry(logrus.StandardLogger())
	fields := logrus.Fields{}
	b.ResetTimer()
	b.ReportAllocs()
	for i := 0; i < b.N; i++ {
		fields[strconv.Itoa(i)] = i
	}
	entry.WithFields(fields).Info("hi")
}

That's actually a good point here; this code is ran on ImageStore.restore(), which iterates over all images (which could be "many"). The logger was added to make the code a bit more DRY (see discussion with @vvoland above #44426 (comment)), but looking now, I don't think that's a good idea for this specific case;

• the logger is only used in error conditions (the code is treats these as non-fatal, but logs them instead)
• but the logger would be constructed unconditionally, which means a lot of allocations for something that (in most cases) wouldn't be used.

Let me change it back to instantiate the logger.WithField(s) only if it's gonna be used. It's less "pretty" but probably good for this case.