ffeac298a2
* fix: ignore unknown fields in more RPC responses * Remove mdbook infrastructure * Delete gitattributes and other theme related items Move all docs to /docs folder to support Docusaurus * all docs need to be moved to /docs * can be changed in the future Add Docusaurus infrastructure * initialize docusaurus repo Remove trailing whitespace, add support for eslint Change Docusaurus configuration to support `src` * No need to rename the folder! Change a setting and we're all good to go. * Fixing rebase items * Remove unneccessary markdown file, fix type * Some fonts are hard to read. Others, not so much. Rubik, you've been sidelined. Roboto, into the limelight! * As much as we all love tutorials, I think we all can navigate around a markdown file. Say goodbye, `mdx.md`. * Setup deployment infrastructure * Move docs job from buildkite to travic * Fix travis config * Add vercel token to travis config * Only deploy docs after merge * Docker rust env * Revert "Docker rust env" This reverts commit f84bc208e807aab1c0d97c7588bbfada1fedfa7c. * Build CLI usage from docker * Pacify shellcheck * Run job on PR and new commits for publication * Update README * Fix svg image building * shellcheck Co-authored-by: Michael Vines <mvines@gmail.com> Co-authored-by: Ryan Shea <rmshea@users.noreply.github.com> Co-authored-by: publish-docs.sh <maintainers@solana.com>
2.5 KiB
title
| title |
|---|
| Validator |
History
When we first started Solana, the goal was to de-risk our TPS claims. We knew that between optimistic concurrency control and sufficiently long leader slots, that PoS consensus was not the biggest risk to TPS. It was GPU-based signature verification, software pipelining and concurrent banking. Thus, the TPU was born. After topping 100k TPS, we split the team into one group working toward 710k TPS and another to flesh out the validator pipeline. Hence, the TVU was born. The current architecture is a consequence of incremental development with that ordering and project priorities. It is not a reflection of what we ever believed was the most technically elegant cross-section of those technologies. In the context of leader rotation, the strong distinction between leading and validating is blurred.
Difference between validating and leading
The fundamental difference between the pipelines is when the PoH is present. In a leader, we process transactions, removing bad ones, and then tag the result with a PoH hash. In the validator, we verify that hash, peel it off, and process the transactions in exactly the same way. The only difference is that if a validator sees a bad transaction, it can't simply remove it like the leader does, because that would cause the PoH hash to change. Instead, it rejects the whole block. The other difference between the pipelines is what happens after banking. The leader broadcasts entries to downstream validators whereas the validator will have already done that in RetransmitStage, which is a confirmation time optimization. The validation pipeline, on the other hand, has one last step. Any time it finishes processing a block, it needs to weigh any forks it's observing, possibly cast a vote, and if so, reset its PoH hash to the block hash it just voted on.
Proposed Design
We unwrap the many abstraction layers and build a single pipeline that can toggle leader mode on whenever the validator's ID shows up in the leader schedule.
Notable changes
- Hoist FetchStage and BroadcastStage out of TPU
- BankForks renamed to Banktree
- TPU moves to new socket-free crate called solana-tpu.
- TPU's BankingStage absorbs ReplayStage
- TVU goes away
- New RepairStage absorbs Shred Fetch Stage and repair requests
- JSON RPC Service is optional - used for debugging. It should instead be part
of a separate
solana-blockstreamerexecutable. - New MulticastStage absorbs retransmit part of RetransmitStage
- MulticastStage downstream of Blockstore