github.com/Finschia/finschia-sdk@v0.48.1/x/slashing/spec/07_tombstone.md (about) 1 <!-- 2 order: 7 3 --> 4 5 # Staking Tombstone 6 7 ## Abstract 8 9 In the current implementation of the `slashing` module, when the consensus engine 10 informs the state machine of a validator's consensus fault, the validator is 11 partially slashed, and put into a "jail period", a period of time in which they 12 are not allowed to rejoin the validator set. However, because of the nature of 13 consensus faults and ABCI, there can be a delay between an infraction occurring, 14 and evidence of the infraction reaching the state machine (this is one of the 15 primary reasons for the existence of the unbonding period). 16 17 > Note: The tombstone concept, only applies to faults that have a delay between 18 > the infraction occurring and evidence reaching the state machine. For example, 19 > evidence of a validator double signing may take a while to reach the state machine 20 > due to unpredictable evidence gossip layer delays and the ability of validators to 21 > selectively reveal double-signatures (e.g. to infrequently-online light clients). 22 > Liveness slashing, on the other hand, is detected immediately as soon as the 23 > infraction occurs, and therefore no slashing period is needed. A validator is 24 > immediately put into jail period, and they cannot commit another liveness fault 25 > until they unjail. In the future, there may be other types of byzantine faults 26 > that have delays (for example, submitting evidence of an invalid proposal as a transaction). 27 > When implemented, it will have to be decided whether these future types of 28 > byzantine faults will result in a tombstoning (and if not, the slash amounts 29 > will not be capped by a slashing period). 30 31 In the current system design, once a validator is put in the jail for a consensus 32 fault, after the `JailPeriod` they are allowed to send a transaction to `unjail` 33 themselves, and thus rejoin the validator set. 34 35 One of the "design desires" of the `slashing` module is that if multiple 36 infractions occur before evidence is executed (and a validator is put in jail), 37 they should only be punished for single worst infraction, but not cumulatively. 38 For example, if the sequence of events is: 39 40 1. Validator A commits Infraction 1 (worth 30% slash) 41 2. Validator A commits Infraction 2 (worth 40% slash) 42 3. Validator A commits Infraction 3 (worth 35% slash) 43 4. Evidence for Infraction 1 reaches state machine (and validator is put in jail) 44 5. Evidence for Infraction 2 reaches state machine 45 6. Evidence for Infraction 3 reaches state machine 46 47 Only Infraction 2 should have its slash take effect, as it is the highest. This 48 is done, so that in the case of the compromise of a validator's consensus key, 49 they will only be punished once, even if the hacker double-signs many blocks. 50 Because, the unjailing has to be done with the validator's operator key, they 51 have a chance to re-secure their consensus key, and then signal that they are 52 ready using their operator key. We call this period during which we track only 53 the max infraction, the "slashing period". 54 55 Once, a validator rejoins by unjailing themselves, we begin a new slashing period; 56 if they commit a new infraction after unjailing, it gets slashed cumulatively on 57 top of the worst infraction from the previous slashing period. 58 59 However, while infractions are grouped based off of the slashing periods, because 60 evidence can be submitted up to an `unbondingPeriod` after the infraction, we 61 still have to allow for evidence to be submitted for previous slashing periods. 62 For example, if the sequence of events is: 63 64 1. Validator A commits Infraction 1 (worth 30% slash) 65 2. Validator A commits Infraction 2 (worth 40% slash) 66 3. Evidence for Infraction 1 reaches state machine (and Validator A is put in jail) 67 4. Validator A unjails 68 69 We are now in a new slashing period, however we still have to keep the door open 70 for the previous infraction, as the evidence for Infraction 2 may still come in. 71 As the number of slashing periods increase, it creates more complexity as we have 72 to keep track of the highest infraction amount for every single slashing period. 73 74 > Note: Currently, according to the `slashing` module spec, a new slashing period 75 > is created every time a validator is unbonded then rebonded. This should probably 76 > be changed to jailed/unjailed. See issue [#3205](https://github.com/cosmos/cosmos-sdk/issues/3205) 77 > for further details. For the remainder of this, I will assume that we only start 78 > a new slashing period when a validator gets unjailed. 79 80 The maximum number of slashing periods is the `len(UnbondingPeriod) / len(JailPeriod)`. 81 The current defaults in Gaia for the `UnbondingPeriod` and `JailPeriod` are 3 weeks 82 and 2 days, respectively. This means there could potentially be up to 11 slashing 83 periods concurrently being tracked per validator. If we set the `JailPeriod >= UnbondingPeriod`, 84 we only have to track 1 slashing period (i.e not have to track slashing periods). 85 86 Currently, in the jail period implementation, once a validator unjails, all of 87 their delegators who are delegated to them (haven't unbonded / redelegated away), 88 stay with them. Given that consensus safety faults are so egregious 89 (way more so than liveness faults), it is probably prudent to have delegators not 90 "auto-rebond" to the validator. 91 92 ### Proposal: infinite jail 93 94 We propose setting the "jail time" for a 95 validator who commits a consensus safety fault, to `infinite` (i.e. a tombstone state). 96 This essentially kicks the validator out of the validator set and does not allow 97 them to re-enter the validator set. All of their delegators (including the operator themselves) 98 have to either unbond or redelegate away. The validator operator can create a new 99 validator if they would like, with a new operator key and consensus key, but they 100 have to "re-earn" their delegations back. 101 102 Implementing the tombstone system and getting rid of the slashing period tracking 103 will make the `slashing` module way simpler, especially because we can remove all 104 of the hooks defined in the `slashing` module consumed by the `staking` module 105 (the `slashing` module still consumes hooks defined in `staking`). 106 107 ### Single slashing amount 108 109 Another optimization that can be made is that if we assume that all ABCI faults 110 for Tendermint consensus are slashed at the same level, we don't have to keep 111 track of "max slash". Once an ABCI fault happens, we don't have to worry about 112 comparing potential future ones to find the max. 113 114 Currently the only Tendermint ABCI fault is: 115 116 - Unjustified precommits (double signs) 117 118 It is currently planned to include the following fault in the near future: 119 120 - Signing a precommit when you're in unbonding phase (needed to make light client bisection safe) 121 122 Given that these faults are both attributable byzantine faults, we will likely 123 want to slash them equally, and thus we can enact the above change. 124 125 > Note: This change may make sense for current Tendermint consensus, but maybe 126 > not for a different consensus algorithm or future versions of Tendermint that 127 > may want to punish at different levels (for example, partial slashing).