Recent advances like P4 programmable hardware switches in the network, and eBPF programs in the endhost network stack, have significantly improved the ability to customize packet processing pipelines in middleboxes and endhosts, and have enabled the offload of some simple application layer processing to the network. These programming frameworks are typically used for parsing mostly fixed-format network or transport layer headers, and are considered unsuitable for parsing complex application layer headers with variable formats. This paper characterizes the feasibility and limits of parsing complex application layer messages within the restrictive programming environments of P4 switches and eBPF kernel programs, a question that has not received much attention in prior work. We evaluate the feasibility of parsing different types of application message formats on the forwarding path using existing parsing techniques and quantify the overhead of such parsing on the application performance using our optimized implementations of the parsing techniques in eBPF and P4. We use on-path application telemetry, where application layer metrics are extracted by parsing application headers on the packet forwarding path, as a case study to evaluate our implementation. Our evaluation shows that, within the limits of feasibility, on-path network telemetry is more efficient and makes the metrics available sooner than state-of-the-art off-path telemetry systems that mirror packets and analyze them in userspace software.