Slashing: special kind of slashing-by-transaction for infractions older than an unbonding period

[cwgoes]

Our slashing model is built on "follow-the-stake": we want to slash stake which contributed to an infraction which is discovered later (after the state has changed), so we need the ability to recover the state of the stake (which accounts contributed how much by delegation or self-bonding) at the time of the infraction. Presently, Gaia only tracks enough information (redelegations / unbonding delegations) to recover the state of a validator's stake for infractions committed an unbonding period or less prior to discovery (an archive node can elect to track all past states, but we probably don't want to require that).

If we want to slash bonded validators for infractions committed longer than an unbonding period in the past - #1022, to encourage key rotation - we could create a special slashing transaction which proved a double-signature at the time of infraction (exposing a way to load previous headers in the handler and charging gas on the transaction to prevent spam), or just have the slashing BeginBlock function accept evidence older than unbonding period (depending on whether we want to support slashing evidence beyond Tendermint's maximum evidence age).

The tricky part is following all the stake. Unbonded or redelegated stake is no longer eligible to be slashed after an unbonding period, but we still need to calculate somehow the changes to the stake since the infraction, so we can avoid slashing stake which has bonded to the validator since the infraction. In principle, the transaction itself could include proofs of stake changes, although this seems complex to implement.

Another option is simply to slash stake which has bonded since the infraction even though it didn't contribute, which removes all this complex calculation.

@ebuchman Is this summary accurate / am I missing an obvious solution here?

[rigelrozanski]

• Tokens that are no longer bonded or unbonding(/redelegating) from the a misbehaving validator can simply not incur slashes for earlier (>unbonding-period) behaviour of a previous validator.
• The premise of cosmos is that is does not meddle with stake which has been unbonded from the violating validator.
• the protocol could however choose to arbitrarily slash-by-association additional stake which did not actually participate in bad behaviour but is bonded to the validator which did - I don't see a problem with this so long as the circumstances are explicitly communicated.

[cwgoes]

Tabled to post-launch in favor of #2081.

[jackzampolin]

Is this still something we want to implement?

[cwgoes]

Is this still something we want to implement?

I think this needs to be part of a broader discussion of light client security & efficiency.

That's where decisions here are most relevant.

[jackzampolin]

Going to link this to the applicable issue on gaia and close it here.