Recent activity in php-src is relevant to teams that run PHP extraction and document processing workers. The window contains 110 commits across 928 files, but the useful signal for data engineers is narrower: clearer HTTP failures, safer XML operations, and tighter behavior in archive and array processing.
HTTP extraction stops at the right failure boundary ¶
A callback exception during a curl transfer is not a recoverable transport detail. It usually means application code rejected a chunk, failed to write a destination, or could not process a response header. The curl callback exception change now aborts the transfer when a callback throws.
That is the correct failure mode for an extractor. Continuing after the consumer has failed can leave a partial object or an incomplete parser state that later code mistakes for valid input. A stopped transfer also gives the scheduler a clean exception boundary for retry policy and failure accounting. The tradeoff is that jobs which accidentally relied on continued transfer activity may now fail sooner, so retry counts and destination cleanup deserve attention.
Before rollout, capture the current number of callback exceptions, partial destination files, and retry attempts for one representative job. After the update, compare those counts with transferred byte totals. A clean failure is useful only if the worker removes incomplete output and exposes the exception to the orchestrator.
Diagnosis also improves. A related change adds the curl option name to error messages. The supported source shows repeated work in ext/curl/interface.c, which is the file to review when downstream builds begin reporting different callback behavior.
Deep XML trees should not take down a worker ¶
XML feeds can be structurally valid and still be hostile to a process stack. Two changes target that exact boundary. One fixes stack overflow in DOMNode::isEqualNode, while another fixes stack overflow during normalization of deeply nested documents.
For ingestion services, a stack overflow is worse than a normal parse error. It can terminate the worker before application cleanup, checkpoint writes, or dead letter routing runs. These patches therefore improve failure containment for feeds that compare or normalize deeply nested nodes. The regression coverage is visible in ext/dom/tests/gh23120.phpt, including a separate adjustment for Windows x86.
The fixes do not make unbounded XML safe. Operators should still cap document size, nesting depth where the parser permits it, and total processing time. Windows builders also have two libxml changes to track: exporting xmlCtxtSetOptions and restoring builds with libxml2 2.15. The related export list is ext/libxml/php_libxml2.def.
Archive selection and set matching get tighter ¶
Archive ingestion often starts with a glob or regular expression that decides which files enter the pipeline. The ZipArchive default options fix says that addGlob() and addPattern() had ignored their default options. That is a correctness issue, not just API polish. A change in selected paths can alter row counts, duplicate handling, and the contents of a generated delivery archive.
Tests should assert the exact archive member list before and after a runtime update. This is especially important when a wrapper omits the options argument because it expects PHP defaults to apply. The implementation area is ext/zip/php_zip.c.
Transform stages also get a targeted performance change. array_intersect() now uses hash based matching. This matters when workers intersect large identifier sets for filtering or reconciliation. No benchmark numbers are present in the captured review, so production gains should be measured with representative key counts and value types rather than assumed.
Poll waits gain a duration type ¶
The new Time\Duration API reaches asynchronous I/O through a change that lets Io\Poll\Context::wait() accept a duration. A follow up corrects errors for negative timeouts.
Typed duration values can make wait budgets clearer than loose numeric conventions. They also create a migration surface for event loops and internal scheduling code. Data services should test zero, negative, and very large waits, plus cancellation behavior around HTTP and database clients. The implementation touched ext/standard/io_poll.c and ext/date/time_duration.c.
What to watch ¶
Read UPGRADING before adopting a build from this branch. It changed eight times in the window, which suggests more compatibility detail than the headline commits alone provide.
Run one malformed XML corpus and one curl callback failure through a disposable worker. Verify exit status, retry classification, temporary file cleanup, and checkpoint state.
Benchmark large
array_intersect()calls with real identifiers. The hash based path is promising, but memory use and type distribution still decide whether it helps a specific transform.