Tags: pinterest/PINCache
Tags
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Fix crashes and unbounded memory in disk cache trimming (#336) * Fix crashes and unbounded memory in disk cache trimming trimToSize:, trimToSizeByEvictionStrategy:, and trimToDate: sorted the entire metadata dictionary with keysSortedByValueUsingComparator:, which materializes O(N) temporary storage inside CoreFoundation. On large caches under memory pressure that allocation fails and traps in __CFCreateArrayStorage — and since every subsequent trim hits the same wall, a cache that grows too big to sort can never shrink again. - Select eviction candidates in bounded 256-key batches (single pass, binary-search insertion, O(batch) memory) and loop until under the byte target, so trims always make progress. An attempted-keys set guarantees termination when eviction declines ghost entries. - trimToDate: is a pure threshold filter; drop its sort entirely. - Nil-safe comparators: metadata dates/size are nil when resourceValuesForKeys: fails during the initial disk scan. [NSNumber compare:nil] RAISES NSInvalidArgumentException (a real crash in trimToSize:), and [NSDate compare:nil] returns NSOrderedSame, which made the sorts weakly inconsistent. Unknown-metadata entries now sort first and are evicted before entries with known access times. - Guard NSUInteger underflow in the trim loops: stale metadata sizes overshooting byteCount made the break condition unsatisfiable and evicted the entire cache. Adds a multi-batch trim regression test (300 entries > one batch). * Fix CI for current macos-latest runner image (macOS 26) CI was failing at the first xcrun/xcodebuild invocation with "xcrun: error: missing DEVELOPER_DIR path: /Applications/Xcode_15.2.app". Root cause: GitHub's macos-latest label now resolves to the macOS 26 runner image, which only ships Xcode 26.x — the pinned Xcode 15.2 (and several other things the build relied on) no longer exist on the image. Changes, following TextureGroup/Texture's CI setup: - DEVELOPER_DIR now points at Xcode_26.5.0.app (the image default; see the runner-images manifest linked in the workflow comment). - Makefile simulator destination bumped from "iPhone 15" to "iPhone 17", since the iPhone 15 simulator does not exist under Xcode 26 on the image. - xcpretty replaced with xcbeautify: the runner images stopped preinstalling the xcpretty gem (macOS 26 ships only xcbeautify 3.2.1), so the "xcodebuild | xcpretty" pipes would fail with command-not-found. Error propagation is unchanged — the Makefile's "SHELL=/bin/bash -o pipefail" still surfaces xcodebuild failures through the pipe. - Removed the platform matrix from the test job; nothing ever referenced matrix.platform (the Makefile owns the destination). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Dan Reed <dreed@pinterest.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
PreviousNext
