Is incremental processing the reason to choose CocoIndex?
CocoIndex is worth evaluating when you need to update derived data repeatedly as its sources change. The README describes a Python interface backed by a Rust engine: declare the desired target state, run an initial backfill, then process changes on later runs. Its embedding example says unchanged files avoid re-embedding. We did not measure that behavior, its savings or its performance at scale.
For a one-off conversion, first decide whether maintaining state between runs solves a problem you actually have. For an index that changes regularly, the useful question is whether updates and deletions reach the target correctly under your workload.
What our recorded run establishes
Our stored run took place on October 3, 2026, at commit e26e9757d10f8c2a7a57754132bf799227297279. No new execution was performed for this correction.
The install step succeeded in 264 seconds, and the build step succeeded in 5 seconds. Recorded dependency metrics were 203 packages and 6,084 MB. The available record does not separate dependency download time from installation work or establish that an installed CLI was checked. It also contains no completed PDF conversion, embedding benchmark or production workload.
The sandbox configuration recorded 3 CPUs, 8 GB of memory and image ghcr.io/astral-sh/uv:python3.12-bookworm. The install log explicitly says it used a Python 3.11.2 environment. Use that runtime observation rather than inferring Python 3.12 from the image name.
The checkout contained 1,187 files, approximately 194,733 source lines and 44.4 MB. Detection found 10 CI workflows and Python tests, with no root tests directory, Dockerfile or Compose file. Those are repository observations, not evidence of connector reliability.
The failed test invocation needs diagnosis
The harness test step lasted 527 seconds and exited with code 1. Its log reported 190 failed, 375 passed, 10 skipped and 10 errors, and said it was stopping after 200 failures. The pytest summary's own elapsed time was 516.95 seconds; that is distinct from the harness step duration.
The parser also stored a total of 575. We have not reconciled the complete log or verified unique, exhaustive coverage, so these figures should not become a pass percentage or a claim about the full suite.
Many displayed core-test failures say that async functions are not natively supported. This identifies an async-test execution limitation in this invocation. It does not establish that the corresponding application behavior is broken. The tail also names errors in Postgres target, SurrealDB target, Turbopuffer target and Valkey target test modules, without enough detail to diagnose each error. Neither missing credentials nor connector defects can be assigned as a universal cause from this tail.
The audit recorded 0 advisories, but its coverage is unknown. We cannot turn that result into a clean dependency-security assessment.
A useful first evaluation
The supplied quickstart is labeled v1.0.14 and last reviewed July 4, 2026. It documents converting local PDFs to Markdown with Docling, configuring COCOINDEX_DB to use ./cocoindex.db, and running cocoindex update main.py. It describes processing added or modified files and removing the corresponding Markdown output after a source deletion. These are documented behaviors, not observations from our sandbox run.
Follow that example before adding your own services. Check the initial output, change one input, delete another and confirm the resulting target state. Then test a transformation change and restart recovery. For an embedding pipeline, record model calls, elapsed time and tracking storage alongside target storage. Those checks would answer workload questions that our setup run cannot.
Read connector reports with their status and scope
The supplied extract for Doris issue 2358 describes a report against aee7b27d in which failed column DDL could leave tracking state ahead of the actual schema. The issue is marked closed as completed in PR 2359 on September 4, 2026. It should not be presented as a currently open defect. Closure also does not substitute for testing the connector and version you deploy; we did not reproduce the report or verify the fix in our run.
When to consider Dagster or dlt
Dagster's README describes an orchestrator for declared data assets with lineage and observability. Consider it when scheduling and coordinating asset-producing functions is the central requirement.
dlt's README describes loading source data into structured, typed datasets, with schema inference, normalization and incremental loading. Consider it when extraction and loading are the main job.
These are comparisons of documented roles. We did not run either alternative, establish feature parity or show that one replaces every part of a CocoIndex pipeline.
How to read the scores
The scores are editorial assessments, not benchmark results. The rubric is: 1 means substantial documented obstacles; 2 means limited supporting evidence; 3 means a workable evaluation starting point with material gaps; 4 means broad supporting evidence with limited gaps; 5 requires comprehensive supporting evidence for the category. Missing evidence lowers confidence and does not prove poor product quality.
Setup receives 3 because the recorded install and build steps succeeded, while test execution needs investigation and no pipeline workload was verified. Documentation receives 3 because the supplied quickstart gives a concrete local workflow but does not establish the quality or completeness of all documentation. Community receives 2 because the supplied material shows participation channels and a completed issue, without establishing support responsiveness or adoption. Maturity receives 3 because the documented workflow and issue resolution support a trial, while this evidence does not establish production reliability.
Begin with one source and one target. Make adoption depend on correct updates, deletions and recovery in that pipeline, plus measurements of the repeated work you want to avoid.